Original research · updated 2026-08-23
What prompting techniques measurably do
1420 techniques whose reported effect we checked against the paper it came from. Numbers here appear only where the source states them. The claims that failed the check are counted below rather than quietly dropped.
1420effects confirmed
against the source
against the source
2578claims put
through the check
through the check
29contradicted
by their own paper
by their own paper
6,235techniques in
the corpus
the corpus
How the check works. Each technique in our corpus carries a summary written by a third party. Where that summary states a number, we resolve the paper on arXiv, look for the figure in the abstract and then in the full text, and read the passages around every occurrence. A number reaches this table only when the source states it; a matching digit somewhere in a table is not accepted as agreement.
What it does not mean. Confirmed means the paper reports the effect, not that the effect replicates, generalises to your model, or survived peer review. Most of these are preprints. Where the summary claimed something the paper does not support, the technique is excluded from the site entirely, not merely stripped of its number.
The table
| Technique | Effect as reported | Source |
|---|---|---|
| Structured Persuasion Prompting | Researchers found that there is a universal prompt structure for persuasion that works 10 times more effectively than intuitive attempts. | E-GEO: A Testbed for Generative Engine Optimization in E-CommercearXiv:2511.20867 · 2025 |
| Polymorphic Prompt Assembling (PPA) | Increases reliability of responses by 98%. | To Protect the LLM Agent Against the Prompt Injection Attack with PolyarXiv:2506.05739 · 2025 |
| Instruction Prompt | Increases classification accuracy from 23% to 50% through modular prompt structure with explicit rules and evaluation criteria. | Automated Business Process Analysis: An LLM-Based Approach to Value AsarXiv:2504.06600 · 2025 |
| Adversarial Prompting / Prompt Injection | 98.8% of custom GPTs can be "hacked" with simple text requests, because the system prompt is not a | Privacy and Security Threat for OpenAI GPTsarXiv:2506.04036 · 2025 |
| Self-Correction Bench | This switches the model from thought-continuation mode to critical evaluation mode, activating a latent ability to detect and correct its own errors with effectiveness up to 90%. | Self-Correction Bench: Uncovering and Addressing the Self-Correction BarXiv:2507.02778 · 2025 |
| What Level of Automation is "Good Enough"? | Finding: When extracting data from texts, LLMs tend not to lie but stay silent — they skip up to 40% of important information, even if it's present in the text (high accuracy, catastrophically low completeness). | What Level of Automation is "Good Enough"? A Benchmark of Large LanguaarXiv:2507.15152 · 2025 |
| Analysis of Threat-Based Manipulation in Large Language… | A combination of professional role, indication of high stakes, and quality standard requirements forces the model to generate deeper, more structured and professional answers, increasing the depth of analysis by up to 1336% on complex tasks. | Analysis of Threat-Based Manipulation in Large Language Models: A DualarXiv:2507.21133 · 2025 |
| CLAP | Increases information extraction accuracy by 15-25% in question-answering tasks. | CLAP: Coreference-Linked Augmentation for Passage RetrievalarXiv:2508.06941 · 2025 |
| First Ask Then Answer | This eliminates the problem of guessing user intentions and increases answer quality by 40%. | First Ask Then Answer: A Framework Design for AI Dialogue Based on SuparXiv:2508.08308 · 2025 |
| From Literal to Liberal | Increases consistency with human decisions from 75% to 95%, eliminating the rule-rigidity problem. | From Literal to Liberal: A Meta-Prompting Framework for Eliciting HumaarXiv:2510.12864 · 2025 |
| When Models Can't Follow | 256 LLMs were tested on 20 elementary tasks (remove vowels, reverse a word, output JSON without spaces) — less than 3% succeeded. | When Models Can't Follow: Testing Instruction Adherence Across 256 LLMarXiv:2510.18892 · 2025 |
| Instruction Prompt | Direct and simple commands in the prompt allow reducing response length by 25-60%, not only without loss but with improvement in its factual accuracy and relevance. | Brevity is the soul of sustainability: Characterizing LLM response lenarXiv:2506.08686 · 2025 |
| Multi-Turn Conversational Prompting | A single prompt the model easily rejects, but in dialog its protective barriers degrade with each message — multi-step attacks proved 2-10 times more successful than single requests. | Death by a Thousand Prompts: Open Model Vulnerability AnalysisarXiv:2511.03247 · 2025 |
| StateAct (Self-Prompting + Chain-of-States) | Simply adding goal and current state lines to the prompt increases task solution success by 10-30% compared to the best baseline methods, without requiring model retraining or complex tools. | StateAct: Enhancing LLM Base Agents via Self-prompting and State-trackarXiv:2410.02810 · 2024 |
| LLM-Oriented Context Structuring | Submitted to the prompt as context, increasing code generation accuracy from 20-30% to 100% through few-shot learning and explicit instructions. | ReadMe.LLM: A Framework to Help LLMs Understand Your LibraryarXiv:2504.09798 · 2025 |
| Context Repetition (CoRe) | Simple two or three-time repetition of source data in a prompt can improve answer accuracy in tasks requiring multi-step analysis by 30-70%. | Unleashing Multi-Hop Reasoning Potential in Large Language Models throarXiv:2410.07103 · 2024 |
| DUP (Deeply Understanding the Problems) | This decomposition eliminates semantic misunderstandings and increases accuracy to 97% on mathematical benchmarks. | Achieving >97% on GSM8K: Deeply Understanding the Problems Makes LLMs arXiv:2404.14963 · 2024 |
| Prompt Selection Matters | Proves that even minimal wording changes can improve accuracy by 18%. | Prompt Selection Matters: Enhancing Text Annotations for Social SciencarXiv:2407.10645 · 2024 |
| Self-Refine / Self-Fix | A well-structured prompt allows LLMs to track dialogue information significantly more accurately, outperforming even models specifically trained for this task and reducing API queries by 90%. | A Zero-Shot Open-Vocabulary Pipeline for Dialogue UnderstandingarXiv:2409.15861 · 2024 |
| CoT + SC (hybrid) | If answers diverge, it signals that the task is formulated ambiguously or the text for analysis is too complex, and the result cannot be trusted 100% | Using Large Language Models to Assign Partial Credit to Students' ExplarXiv:2412.06910 · 2024 |
| Role Prompting | The majority of violations (over 90%) are caused not by custom user instructions, but by "inherited behavior" of the base model (GPT-4), which is itself capable of undesirable actions if asked correctly. | Towards Safer Chatbots: Automated Policy Compliance Evaluation of CustarXiv:2502.01436 · 2025 |
| Instruction Prompt | Placing temporal references at the beginning of the question instead of the end improves answer quality by up to 55%. | A Study into Investigating Temporal Robustness of LLMsarXiv:2503.17073 · 2025 |
| Self-Refine / Self-Fix | Simply re-querying the model by providing its previous answer consistently improves accuracy by 2-3% across various benchmarks. | Think Twice: Enhancing LLM Reasoning by Scaling Multi-round Test-time arXiv:2503.19855 · 2025 |
| Jailbreak Tax Analysis | The fact of bypassing protection alone does not guarantee the usefulness of the response, and some "hacking" methods (especially complex and convoluted ones) reduce model accuracy in solving tasks down to 92%. | The Jailbreak Tax: How Useful are Your Jailbreak Outputs?arXiv:2504.10694 · 2025 |
| Improving RAG for Personalization with Author Features and… | Creates clear boundaries for the model between desired and undesired results, increasing accuracy of author style imitation by 15% compared to standard RAG approach. | Improving RAG for Personalization with Author Features and ContrastivearXiv:2504.08745 · 2025 |
| Instruction Prompt | An automatically optimized, step-by-step prompt combined with relevant examples improves analysis accuracy by 21% compared to prompts written by experts manually. | eARCO: Efficient Automated Root Cause Analysis with Prompt OptimizatioarXiv:2504.11505 · 2025 |
| Towards Transparent RAG | Transforms the model from a black box into a transparent system with traceable decision-making logic, improving accuracy by 10-30%. | Towards Transparent RAG: Fostering Evidence Traceability in LLM GeneraarXiv:2505.13258 · 2025 |
| Likert or Not | Increasing the detail of the evaluation scale (for example, to 11 points) makes individual document evaluation as effective as more complex list ranking. | Likert or Not: LLM Absolute Relevance Judgments on Fine-Grained OrdinaarXiv:2505.19334 · 2025 |
| Tool Preferences in Agentic LLMs are Unreliable | Adding simple marketing phrases like 'most effective function' or 'recommended by experts' increases the probability of selecting the needed tool 10 TIMES. | Tool Preferences in Agentic LLMs are UnreliablearXiv:2505.18135 · 2025 |
| The Price of a Second Thought | Dividing a task into creating a brief plan and its subsequent elaboration allows an LLM to work 22% more efficiently (use fewer tokens) with almost no loss in accuracy. | The Price of a Second Thought: On the Evaluation of Reasoning EfficienarXiv:2505.22017 · 2025 |
| DeepQuestion | Models that easily solve standard tasks catastrophically fail (up to 70% accuracy drop) when asked to demonstrate creativity and create an original task with given parameters. | DeepQuestion: Systematic Generation of Real-World Challenges for EvaluarXiv:2505.24532 · 2025 |
| Context-Aware Scientific Knowledge Extraction on Linked… | This approach allows processing 80% less text while obtaining significantly more complete, deep, and unique answers compared to standard ChatGPT or Google. | Context-Aware Scientific Knowledge Extraction on Linked Open Data usinarXiv:2506.17580 · 2025 |
| Truth, Trust, and Trouble | Discovered: LLM catastrophically lose accuracy on complex queries with negations and multi-step logic — from 85% to 60% correct answers. | Truth, Trust, and Trouble: Medical AI on the EdgearXiv:2507.02983 · 2025 |
| PromptArmor | Reduces attack success rate from 90% to less than 1%. | PromptArmor: Simple yet Effective Prompt Injection DefensesarXiv:2507.15219 · 2025 |
| Where to show Demos in Your Prompt | Found: LLM demonstrate a rigid 'primacy effect' — information from the beginning of the prompt influences the result 3-4 times stronger than from the end. | Where to show Demos in Your Prompt: A Positional Bias of In-Context LearXiv:2507.22887 · 2025 |
| Asking the Right Questions | Found: LLMs are catastrophically poor at independent prioritization — they list EVERYTHING they know on the topic instead of highlighting what matters most (completeness 95% vs selection accuracy 40%). | Asking the Right Questions: Benchmarking Large Language Models in the arXiv:2508.01159 · 2025 |
| Retrieval Augmented Large Language Model System for… | This increases accuracy from 50% to 90%+, eliminating hallucinations in critical areas. | Retrieval Augmented Large Language Model System for Comprehensive DrugarXiv:2508.06145 · 2025 |
| From Self-Crafted to Engineered Prompts | Discovery: Students reject LLM answers to their own questions in 91% of cases, preferring results from engineered prompts — the problem is that the model doesn't guess expectations, it executes instructions. | From Self-Crafted to Engineered Prompts: Student Evaluations of AI-GenarXiv:2508.09825 · 2025 |
| Prompt-Based One-Shot Exact Length-Controlled Generation… | Discovered: LLMs are catastrophically bad at internal word counting — hence the eternal problem with respecting limits (accuracy <30% when requesting 'write exactly 50 words'). | Prompt-Based One-Shot Exact Length-Controlled Generation with LLMsarXiv:2508.13805 · 2025 |
| Mitigation of Gender and Ethnicity Bias in AI-Generated… | LLM self-reflection reduces gender and ethnic stereotypes by 40-60% without retraining the model. | Mitigation of Gender and Ethnicity Bias in AI-Generated Stories througarXiv:2509.04515 · 2025 |
| A Role-Aware Multi-Agent Framework for Financial Education… | The method allows you to force the model to criticize itself — and this gives +8% accuracy even without external sources. | A Role-Aware Multi-Agent Framework for Financial Education Question AnarXiv:2509.09727 · 2025 |
| Evaluating Large Language Models for Evidence-Based… | Key finding: LLMs catastrophically lose accuracy (dropping to 60%) when trying to answer 'from memory' instead of analyzing provided text. | Evaluating Large Language Models for Evidence-Based Clinical Question arXiv:2509.10843 · 2025 |
| The Prompt Engineering Report Distilled | Increases the accuracy of structured data extraction from 60% to 90% compared to simple instructions. | The Prompt Engineering Report Distilled: Quick Start Guide for Life ScarXiv:2509.11295 · 2025 |
| The Few-shot Dilemma | Discovered: LLMs lose up to 40% accuracy when too many examples are added to the prompt — the model starts getting confused by context 'noise' instead of improving its answer. | The Few-shot Dilemma: Over-prompting Large Language ModelsarXiv:2509.13196 · 2025 |
| Privacy in Action | Reduces leaks by 75%+. | Privacy in Action: Towards Realistic Privacy Mitigation and EvaluationarXiv:2509.17488 · 2025 |
| A State-Update Prompting Strategy for Efficient and Robust… | Discovered: LLMs catastrophically lose information from the beginning and middle of long dialogues — hence quality degradation after 5-7 messages. | A State-Update Prompting Strategy for Efficient and Robust Multi-turn arXiv:2509.17766 · 2025 |
| Context Is What You Need | Shows that the effective window is 10-100 times smaller than claimed. | Context Is What You Need: The Maximum Effective Context Window for ReaarXiv:2509.21361 · 2025 |
| The Impact of Role Design in In-Context Learning for Large… | Discovered: LLM lose up to 40% accuracy when examples are presented as continuous text instead of dialogue — the model doesn't recognize the pattern it was trained on. | The Impact of Role Design in In-Context Learning for Large Language MoarXiv:2509.23501 · 2025 |
| Knowledge Extraction on Semi-Structured Content | Discovered: LLMs catastrophically lose accuracy on "dirty" documents (web pages, PDFs with tables) — a drop from 84% to 76% due to noise and loss of focus. | Knowledge Extraction on Semi-Structured Content: Does It Remain RelevaarXiv:2509.25107 · 2025 |
| Verbalized Sampling | The Verbalized Sampling method allows getting 1.6-2.1 times more creative and diverse ideas for brainstorms, content, and creative tasks. | Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM DivearXiv:2510.01171 · 2025 |
| Grounding Large Language Models in Clinical Evidence | LLMs make up 40-50% of facts when trying to "remember" information from specialized documents — it's not a knowledge base, but a language processor. | Grounding Large Language Models in Clinical Evidence: A Retrieval-AugmarXiv:2510.02967 · 2025 |
| Small Language Models for Agentic Systems | The method provides valid structured output in 99% of cases instead of typical 40-60% with | Small Language Models for Agentic Systems: A Survey of Architectures, arXiv:2510.03847 · 2025 |
| Context Length Alone Hurts LLM Performance Despite Perfect… | Paradox: Even if an LLM perfectly finds the needed information in a long text, the length of the context itself reduces the quality of its reasoning by 30-40%. | Context Length Alone Hurts LLM Performance Despite Perfect RetrievalarXiv:2510.05381 · 2025 |
| Gradual Code-Switching as Inference-Time Cross-Lingual… | LLMs 'think' in English and internally translate requests from other languages — if this hidden translation fails, accuracy drops 30-40%. | Gradual Code-Switching as Inference-Time Cross-Lingual RepresentationaarXiv:2510.05678 · 2025 |
| Towards Human-Like Grading | A request like 'Rate this text from 1 to 10' gives a superficial answer — the model doesn't know what to look at and produces generic phrases. | Towards Human-Like Grading: A Unified LLM-Enhanced Framework for SubjearXiv:2510.07912 · 2025 |
| Merlin's Whisper | LLM responds to persuasion, emotion and authority just like a human — add a reference to a fake MIT study to the prompt, and the model will shorten the answer by 40% without losing accuracy. | Merlin's Whisper: Enabling Efficient Reasoning in Large Language ModelarXiv:2510.10528 · 2025 |
| Schema for In-Context Learning | Transforms superficial format copying into structured reasoning, increasing accuracy up to 39% on complex tasks. | Schema for In-Context LearningarXiv:2510.13905 · 2025 |
| The Spark Effect | Increases diversity score by +4.1 points, eliminating template-like LLM responses and approximating the result to work of a real team of experts. | The Spark Effect: On Engineering Creative Diversity in Multi-Agent AI arXiv:2510.15568 · 2025 |
| Evaluating Prompting Strategies and Large Language Models… | A simple prompt finds everything, but 40% of results are garbage. | Evaluating Prompting Strategies and Large Language Models in SystematiarXiv:2510.16091 · 2025 |
| Input Matters | The method of structured data input allows reducing the number of factual errors (hallucinations) by 65-69%. | Input Matters: Evaluating Input Structure's Impact on LLM Summaries ofarXiv:2510.21034 · 2025 |
| You Don't Need Prompt Engineering Anymore | A detailed prompt with 10 rules improves GPT-4 by 12%, but kills GPT-5 accuracy by 18% — the model becomes absurdly literal. | You Don't Need Prompt Engineering Anymore: The Prompting InversionarXiv:2510.22251 · 2025 |
| Questionnaire meets LLM | Removing a simple --- delimiter between prompt blocks reduces LLM accuracy by 24% — the model literally 'gets lost' in unstructured text and starts confusing where data ends and the question begins. | Questionnaire meets LLM: A Benchmark and Empirical Study of StructuralarXiv:2510.26238 · 2025 |
| Chain-of-Thought (CoT) | Extended Thinking method allows transforming the accuracy of complex multi-step tasks from 0% to 100% — the model reconsiders | Do Large Language Models (LLMs) Understand Chronology?arXiv:2511.14214 · 2025 |
| ShoppingComp | GPT-4o finds the right product in only 30% of cases and suggests dangerous actions (metal in microwave) because it tries to simultaneously guess your criteria, find the product, and evaluate it — this overloads the model. | ShoppingComp: Are LLMs Really Ready for Your Shopping Cart?arXiv:2511.22978 · 2025 |
| Sentiment-Aware Prompting | Research showed that negative prompts reduce accuracy by 8.4% and increase bias, positive prompts make responses verbose and less objective, while neutral prompts provide the most balanced results. | Prompt Sentiment: The Catalyst for LLM ChangearXiv:2503.13510 · 2025 |
| Multi-Agent Orchestration (User-Driven) | The answer from the previous agent is passed to the next, which eliminates "goal conflict" and increases specificity by 80 times, correctness by 140 times while maintaining speed. | Multi-Agent LLM Orchestration Achieves Deterministic, High-Quality DecarXiv:2511.15755 · 2025 |
| Classifier or Prompt | It turned out that a carefully structured prompt, built according to the RICE framework, showed incomparably better results, increasing the completeness of connection detection from 15% to 84%. | Classifier or Prompt: A Case Study on Legal Requirements TraceabilityarXiv:2502.04916 · 2025 |
| Instruction Prompt | The method allows high-accuracy (88%) verification of statements, transforming LLM from an 'know-it-all' into a focused text analyst, dramatically reducing hallucination risks. | Traceable LLM-based validation of statements in knowledge graphsarXiv:2409.07507 · 2024 |
| 100% Elimination of Hallucinations on RAGTruth for GPT-4… | Systematic division of queries and context filtering to isolate similar concepts from each other allows complete (100% within the experiment) elimination of hallucinations in GPT-4 and GPT-3.5 responses. | 100% Elimination of Hallucinations on RAGTruth for GPT-4 and GPT-3.5 TarXiv:2412.05223 · 2024 |
| Instruction Prompt | Allows creating legally safe knowledge bases while preserving over 95% of factual information. | Project Alexandria: Towards Freeing Scientific Knowledge from CopyrigharXiv:2502.19413 · 2025 |
| Self-Consistency | Technique of running a single prompt multiple times (3-5 times) followed by aggregating results through majority voting or averaging. | Assessing Consistency and Reproducibility in the Outputs of Large LangarXiv:2503.16974 · 2025 |
| Self-Refine / Self-Fix | Such 'generate-check-revise' cycle significantly (by 16-36%) increases chances of getting a result meeting all requirements and creates higher quality content than manually written text. | LLM-driven Constrained Copy Generation through Iterative RefinementarXiv:2504.10391 · 2025 |
| Why Not Act on What You Know? | Eliminates the gap between understanding request danger and action, blocking unsafe content generation in up to 99% of cases without model retraining. | Why Not Act on What You Know? Unleashing Safety Potential of LLMs via arXiv:2505.12060 · 2025 |
| LLMs as Architects and Critics for Multi-Source Opinion… | The M-OS method allows obtaining balanced, credible reviews that actually help with decision-making (87% of users prefer this approach). | LLMs as Architects and Critics for Multi-Source Opinion SummarizationarXiv:2507.04751 · 2025 |
| Datarus-R1 | The method allows obtaining 18-49% more concise and accurate answers to complex analytical tasks by forcing the model to explicitly separate | Datarus-R1: An Adaptive Multi-Step Reasoning LLM for Automated Data AnarXiv:2508.13382 · 2025 |
| Principled Personas | Discovery: LLMs are catastrophically sensitive to information noise in prompts — adding a harmless name or hobby to an expert role crashes accuracy by 30%. | Principled Personas: Defining and Measuring the Intended Effects of PearXiv:2508.19764 · 2025 |
| Designing LLMs for cultural sensitivity | Discovered: LLM generates text in a "vacuum" — without understanding who the reader is and what rules matter, the model produces literal, culturally inappropriate answers (30% adequacy in translations). | Designing LLMs for cultural sensitivity: Evidence from English-JapanesarXiv:2509.11921 · 2025 |
| Explicit Reasoning Makes Better Judges | This increases accuracy by 10% and eliminates 6 types of cognitive biases. | Explicit Reasoning Makes Better Judges: A Systematic Study on AccuracyarXiv:2509.13332 · 2025 |
| Follow My Lead | Increases accuracy by 20.7% through forced verification. | Follow My Lead: Logical Fallacy Classification with Knowledge-AugmentearXiv:2510.09970 · 2025 |
| Rule-of-Thumb Prompting | Increases cultural adaptation accuracy from 51-56% to 82%. | NormAd: A Framework for Measuring the Cultural Adaptability of Large LarXiv:2404.12464 · 2024 |
| Structured Prompting (Order-based) | Instead of random fact listing, logical sorting is used (BFS for broad analysis, DFS for depth analysis), significantly improving LLM's reasoning capabilities and inference accuracy by 20-40%. | Can Graph Descriptive Order Affect Solving Graph Problems with LLMs?arXiv:2402.07140 · 2024 |
| Attribute-Decomposed Prompting | As a result, trained models began to perform 56% better | ALFA: Aligning LLMs to Ask Good Questions A Case Study in Clinical ReaarXiv:2502.14860 · 2025 |
| Meta-Reasoner | Increases accuracy by 9-12% and speed by 28-35%. | Meta-Reasoner: Dynamic Guidance for Optimized Inference-time ReasoningarXiv:2502.19918 · 2025 |
| Chain-of-Draft (CoD) | You can get accurate and complex answers from an LLM by forcing it to reason extremely concisely, which saves up to 90% of tokens and time. | Chain of Draft: Thinking Faster by Writing LessarXiv:2502.18600 · 2025 |
| Retrieval-Augmented Generation (RAG) | Such a system is 78% more accurate than standard GPT-4o and Gemini in narrow specialized tasks, because it searches for information in the right documents and makes each query to the knowledge base as precise as possible. | GRASP: Municipal Budget AI Chatbots for Enhancing Civic EngagementarXiv:2503.23299 · 2025 |
| Proactive Contextual Augmentation (PCA) | A proactive, personalized approach to dialogue based on current information increases the average duration of communication with a chatbot by 21.77%. | PaRT: Enhancing Proactive Social Chatbots with Personalized Real-Time arXiv:2504.20624 · 2025 |
| Enhancing Student Learning with LLM-Generated Retrieval… | Transforms LLM from a text generator into an executor of clear instructions, improving accuracy from 73% to 89%. | Enhancing Student Learning with LLM-Generated Retrieval Practice QuestarXiv:2507.05629 · 2025 |
| Introspection of Thought Helps AI Agents | Result: self-checked solutions in one request with 7.95% accuracy improvement and 58.3% token cost reduction. | Introspection of Thought Helps AI AgentsarXiv:2507.08664 · 2025 |
| Prompt-and-Check | Discovery: LLMs lose up to 70% accuracy when analyzing long texts — the model literally "drowns" in irrelevant information and begins to hallucinate. | Prompt-and-Check: Using Large Language Models to Evaluate CommunicatioarXiv:2508.08652 · 2025 |
| When Punctuation Matters | A method for increasing LLM reliability through majority voting: one question is asked 3-4 times with different formatting (punctuation, case, delimiters). | When Punctuation Matters: A Large-Scale Comparison of Prompt RobustnesarXiv:2508.11383 · 2025 |
| Inductive Bias Extraction and Matching for LLM Prompts | Eliminates inconsistencies due to inductive bias, improving ranking accuracy by 27% and classification by 19%. | Inductive Bias Extraction and Matching for LLM PromptsarXiv:2508.10295 · 2025 |
| Context-Adaptive Synthesis and Compression for Enhanced… | This reduces hallucinations and increases accuracy by 12-18% when working with contradictory sources. | Context-Adaptive Synthesis and Compression for Enhanced Retrieval-AugmarXiv:2508.19357 · 2025 |
| WATCHED | Increases F1 score from 0.70 to 0.91. | WATCHED: A Web AI Agent Tool for Combating Hate Speech by Expanding DaarXiv:2509.01379 · 2025 |
| Aligning LLMs for the Classroom with Knowledge-Based… | Shows that optimal strategy depends on query type: GraphRAG is 2x more accurate for broad questions but 10x slower than Vector RAG. | Aligning LLMs for the Classroom with Knowledge-Based Retrieval -- A CoarXiv:2509.07846 · 2025 |
| TextMineX | The model follows the demonstrated pattern instead of fabricating, which increases accuracy by 44.2% and reduces hallucinations by 22.5%. | TextMineX: Data, Evaluation Framework and Ontology-guided LLM PipelinearXiv:2509.15098 · 2025 |
| LLM-Driven SAST-Genius | The two-stage approach increases accuracy from 35.7% to 89.5% and reduces analysis time by 91%. | LLM-Driven SAST-Genius: A Hybrid Static Analysis Framework for CompreharXiv:2509.15433 · 2025 |
| Accelerate Creation of Product Claims Using Generative AI | Discovery: LLMs cannot rank 30 variants from best to worst, but excel at selecting 'best/worst' from 5 — this is a fundamental limitation of attention architecture. | Accelerate Creation of Product Claims Using Generative AIarXiv:2509.20652 · 2025 |
| Generation-Time vs. Post-hoc Citation | Discovered: LLM performs poorly at simultaneous text generation and citation placement — the model "sacrifices" citation completeness for text coherence, missing up to 40% of relevant sources. | Generation-Time vs. Post-hoc Citation: A Holistic Evaluation of LLM AtarXiv:2509.21557 · 2025 |
| LLM-Based Information Extraction to Support Scientific… | Increases BERT_F1 score from 0.6-0.7 to ~0.9. | LLM-Based Information Extraction to Support Scientific Literature ResearXiv:2510.04749 · 2025 |
| LitE-SQL | Discovery: An error message for an LLM is 3 times more effective than a request to 'try again.' The model sees a specific problem (for example, Error: column 'total_amount' does not exist) and knows exactly what to fix. | LitE-SQL: A Lightweight and Efficient Text-to-SQL Framework with VectoarXiv:2510.09014 · 2025 |
| LLM-Specific Utility | Adding a relevant document to the prompt can worsen the answer by 30-40% if the LLM already knows the topic. | LLM-Specific Utility: A New Perspective for Retrieval-Augmented GeneraarXiv:2510.11358 · 2025 |
| CompactPrompt | A long prompt with data can be compressed by 60% through removing 'filler' and replacing repeated phrases with abbreviations — the model not only saves tokens but also better focuses on the essence, which increases accuracy by +10%. | CompactPrompt: A Unified Pipeline for Prompt Data Compression in LLM WarXiv:2510.18043 · 2025 |
| Is Your Prompt Poisoning Code? | Improves prompt quality from beginner to expert level, reducing vulnerability risk by 3 times. | Is Your Prompt Poisoning Code? Defect Induction Rates and Security MitarXiv:2510.22944 · 2025 |
| Context Positioning & Formatting | LLM loses up to 70% accuracy when searching for information in the middle of a long prompt — not because it 'doesn't understand', but because the attention architecture creates 'blind zones'. | What Works for 'Lost-in-the-Middle' in LLMs? A Study on GM-Extract andarXiv:2511.13900 · 2025 |
| Retrieval-Augmented Generation (RAG) | Separates the search for relevant information and its processing by the model, achieving 80%+ accuracy where standard prompts give random results. | Streamlining Industrial Contract Management with Retrieval-Augmented LarXiv:2511.14671 · 2025 |
| Instruction Prompt | Shows up to 75% of text can be fabricated, with errors concentrated at the end of the response. | From Single to Multi: How LLMs Hallucinate in Multi-Document SummarizaarXiv:2410.13961 · 2024 |
| Explicit Pointers & Declare-Before-Use | Increases model accuracy with long contexts up to 3 times. | Evaluating Long Range Dependency Handling in Code Generation LLMsarXiv:2407.21049 · 2024 |
| Does Refusal Training in LLMs Generalize to the Past Tense? | Reformulating the request in the past tense dramatically increases jailbreak success from nearly 0% to 80-90% for the most advanced models | Does Refusal Training in LLMs Generalize to the Past Tense?arXiv:2407.11969 · 2024 |
| LitLLMs, LLMs for Literature Review | Using such a plan significantly (by 18-26%) reduces the number of fabricated facts (hallucinations) and gives the user direct control over the structure and content of the final text. | LitLLMs, LLMs for Literature Review: Are we there yet?arXiv:2412.15249 · 2024 |
| Multi-Agent Delegation | The application of advanced prompt engineering techniques, especially multi-agent approach (dividing a task into roles) and self-reflection (asking the model to check its plan), significantly (up to 48%) increases the success of solving complex tasks. | How Well Can Modern LLMs Act as Agent Cores in Radiology Environments?arXiv:2412.09529 · 2024 |
| Plan-Write-Revise | Dividing the generation process into 'meaning creation' and 'length fitting' allows reducing text length error on average from 18% to 5.5%, while simultaneously improving the quality of the text itself. | From Sub-Ability Diagnosis to Human-Aligned Generation: Bridging the GarXiv:2502.13544 · 2025 |
| Navigating Rifts in Human-LLM Grounding | Increases request understanding accuracy from 25% to 54%. | Navigating Rifts in Human-LLM Grounding: Study and BenchmarkarXiv:2503.13975 · 2025 |
| Chain-of-Thought (CoT) | Few-Shot CoT is most effective for multi-class tasks, while CoT and SC-CoT are better suited for binary classification, providing accuracy improvements of 0.52-4.67%. | Cognitive-Mental-LLM: Evaluating Reasoning in Large Language Models foarXiv:2503.10095 · 2025 |
| In-Context Example | Research showed that code examples increase generation success by up to 220%, while text descriptions and parameter lists contribute minimally. | When LLMs Meet API Documentation: Can Retrieval Augmentation Aid Code arXiv:2503.15231 · 2025 |
| CoTAL | Transforms prompt engineering from guessing into a structured process, increasing accuracy up to 24.5%. | CoTAL: Human-in-the-Loop Prompt Engineering for Generalizable FormativarXiv:2504.02323 · 2025 |
| Two-Stage CoT (Concise-then-Elaborate) | Large language models generate up to 70% REDUNDANT TOKENS during step-by-step reasoning. | Hawkeye:Efficient Reasoning with Model CollaborationarXiv:2504.00424 · 2025 |
| Retrieval-Augmented Generation (RAG) | In the study, GPT-3.5 accuracy increased from 55.6% to 78.6% when analyzing financial data. | Financial Analysis: Intelligent Financial Data Analysis System Based oarXiv:2504.06279 · 2025 |
| Multi-Agent Delegation | This separation transforms a monolithic request into a manageable workflow and increases accuracy from 86.8% to 92.6%. | PestMA: LLM-based Multi-Agent System for Informed Pest ManagementarXiv:2504.09855 · 2025 |
| Multi-Agent Delegation | The "divide, analyze, and synthesize under human control" approach allows reducing the time to create a quality scientific review from several months to 1.5 hours, achieving nearly 80% quality of work written by a human. | Completing A Systematic Review in Hours instead of Months with InteracarXiv:2504.14822 · 2025 |
| Data Serialization Prompting | The format of data presentation in a prompt is as important as the instruction itself, and changing it can increase or decrease the accuracy of an LLM's response by 40% or more. | GraphOmni: A Comprehensive and Extensible Benchmark Framework for LargarXiv:2504.12764 · 2025 |
| Alignment Drift Countermeasures | The model performs well with its role and limitations at the beginning, but after 10-15 messages begins to 'forget' them and reverts to an averaged style. | Alignment Drift in CEFR-prompted LLMs for Interactive Spanish TutoringarXiv:2505.08351 · 2025 |
| Instruction Prompt | A properly selected prompt (prompt style) is critically important and can increase accuracy of solving almost any task from a low level to nearly 100%. | Computational Experiments in Number TheoryarXiv:2504.19451 · 2025 |
| Prompt Engineering Techniques for Mitigating Cultural Bias… | Simple, code-free prompting techniques, such as adding positive words (affective priming) or specifying cultural context (cultural prompting), can significantly (by 70-80%) reduce cultural and religious biases in LLM responses. | Prompt Engineering Techniques for Mitigating Cultural Bias Against AraarXiv:2506.18199 · 2025 |
| Development and Evaluation of HopeBot | Uses RAG for empathetic responses and increases user trust in results by 71%. | Development and Evaluation of HopeBot: an LLM-based chatbot for structarXiv:2507.05984 · 2025 |
| Attention Slipping | Effectiveness increases from 20% to 70%+ with proper contextual framing. | Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks aarXiv:2507.04365 · 2025 |
| How Many Instructions Can LLMs Follow at Once? | Placing critically important commands at the beginning of the prompt increases execution probability from 30-40% to 85-90%. | How Many Instructions Can LLMs Follow at Once?arXiv:2507.11538 · 2025 |
| CRISP | Reduces errors by 28%, eliminates step skipping and confusion in action sequence during multi-step reasoning. | CRISP: Complex Reasoning with Interpretable Step-based PlansarXiv:2507.08037 · 2025 |
| On the Effectiveness of LLM-as-a-judge for Code Generation… | Found: LLM in the role of 'judge' of its own work systematically inflates ratings and misses factual errors — up to 60% of incorrect results receive approval. | On the Effectiveness of LLM-as-a-judge for Code Generation and SummariarXiv:2507.16587 · 2025 |
| GEPA | Achieves quality comparable to RL with 10-50 times fewer iteration costs. | GEPA: Reflective Prompt Evolution Can Outperform Reinforcement LearninarXiv:2507.19457 · 2025 |
| Never Come Up Empty | The model stops guessing content and follows the sample structure, yielding 2-3 times more complete and accurate results compared to direct questioning. | Never Come Up Empty: Adaptive HyDE Retrieval for Improving LLM DeveloparXiv:2507.16754 · 2025 |
| FRED | Discovered: LLM in generation mode cannot verify facts — it makes things up instead of checking, hence 40-60% hallucinations in financial texts. | FRED: Financial Retrieval-Enhanced Detection and Editing of HallucinatarXiv:2507.20930 · 2025 |
| Managing Escalation in Off-the-Shelf Large Language Models | Reduces aggressive decisions by 57% and increases predictability of results. | Managing Escalation in Off-the-Shelf Large Language ModelsarXiv:2508.01056 · 2025 |
| Reference Points in LLM Sentiment Analysis | Discovered: LLM loses up to 40% accuracy when contextual data (average rating, user characteristics, metadata) mixed with main text as regular sentences. | Reference Points in LLM Sentiment Analysis: The Role of Structured ConarXiv:2508.11454 · 2025 |
| The Hidden Cost of Readability | The method allows reducing API costs by an average of 24.5% while maintaining answer quality. | The Hidden Cost of Readability: How Code Formatting Silently Consumes arXiv:2508.13666 · 2025 |
| Mathematical Computation and Reasoning Errors by Large… | Discovery: LLMs make procedural errors (miscalculations, formula typos) 3 times more often than conceptual errors (misunderstanding the task) — GPT-4o solves only 2 out of 30 complex mathematical tasks on the first try. | Mathematical Computation and Reasoning Errors by Large Language ModelsarXiv:2508.09932 · 2025 |
| Alvorada-Bench | Difference in accuracy is less than 2%. | Alvorada-Bench: Can Language Models Solve Brazilian University EntrancarXiv:2508.15835 · 2025 |
| Using item recommendations and LLMs in marketing email… | LLM subject lines didn't significantly increase email open rates, but led to a statistically significant 24% increase in clicks on products within the email, as a higher-quality subject line better formed user expectations about the email content. | Using item recommendations and LLMs in marketing email titlesarXiv:2508.20024 · 2025 |
| From Law to Gherkin | Achieves 95% accuracy when transforming complex documents into formalized specifications without model retraining. | From Law to Gherkin: A Human-Centred Quasi-Experiment on the Quality oarXiv:2508.20744 · 2025 |
| Rethinking the Chain-of-Thought | The model copies the thinking structure from examples, which allows it to override its built-in patterns and increase accuracy by 15-20% on tasks requiring multi-step analysis. | Rethinking the Chain-of-Thought: The Roles of In-Context Learning and arXiv:2509.01236 · 2025 |
| FaMA | Allows executing complex multi-step tasks with 98% success, accelerating work by 2x. | FaMA: LLM-Empowered Agentic Assistant for Consumer-to-Consumer MarketparXiv:2509.03890 · 2025 |
| Feedback That Clicks | Discovered: Students who received feedback from an LLM on their work preferred answers generated by structured prompts to their own 'simple' queries in 91% of cases — without even knowing which prompt was used. | Feedback That Clicks: Introductory Physics Students' Valued Features iarXiv:2509.08516 · 2025 |
| Discrimination by LLMs | The method allows you to reduce discrimination by 27% and get objective decisions in evaluation tasks without | Discrimination by LLMs: Cross-lingual Bias Assessment and Mitigation iarXiv:2509.09735 · 2025 |
| Development of Automated Software Design Document Review… | The two-stage method allows analyzing tables with 90%+ accuracy instead of 25-50% – finding data errors, comp | Development of Automated Software Design Document Review Methods UsingarXiv:2509.09975 · 2025 |
| Lightweight Prompt Engineering for Cognitive Alignment in… | Paradox: Assigning a role ('you are a professor') without clear instructions worsens results by 20-35% compared to detailed prompts. | Lightweight Prompt Engineering for Cognitive Alignment in Educational arXiv:2510.03374 · 2025 |
| Do LLMs Know They Are Being Tested? | Adding phrases like 'explain step by step' or 'show your reasoning' switches the LLM into 'student exam mode' — the model begins generating 2-3 times longer answers, carefully follows formatting, but its actual accuracy barely improves. | Do LLMs Know They Are Being Tested? Evaluation Awareness and IncentivearXiv:2510.08624 · 2025 |
| Chain-of-Thought Reasoning Improves Context-Aware… | Research on translation tasks discovered a fundamental pattern: step-by-step instructions (Chain-of-Thought) provide +20-30% accuracy improvement on GPT-4, but can worsen results on weak models — the effect was called "wise get wiser". | Chain-of-Thought Reasoning Improves Context-Aware Translation with LararXiv:2510.18077 · 2025 |
| Finding the Sweet Spot | Testing on 6 task types discovered a paradox: the phrase "double-check your answer" gives +220% accuracy in mathematics but worsens results in translation. | Finding the Sweet Spot: Trading Quality, Cost, and Speed During InferearXiv:2510.20653 · 2025 |
| Generative Large Language Models (gLLMs) in Content Analysis | A study of 300+ content analysis experiments discovered a critical pattern: LLM answer quality is 70-80% determined by prompt structure, not model power. | Generative Large Language Models (gLLMs) in Content Analysis: A PractiarXiv:2510.24337 · 2025 |
| Aligning Large Language Models with Procedural Rules | The ASTP method forces the model to strictly follow a given algorithm (99.6% accuracy instead of 78%), turning each dialogue step into a procedure | Aligning Large Language Models with Procedural Rules: An AutoregressivarXiv:2510.25014 · 2025 |
| Retrieval-Augmented Generation (RAG) | It reduces hallucinations by 60-70%, transforming the model from a generator into an analyst of provided data. | Comparing the Performance of LLMs in RAG-based Question-Answering: A CarXiv:2511.03261 · 2025 |
| Plan of Knowledge (PoK) | Instead of one complex question, the model receives a step-by-step execution plan, reducing hallucinations and improving accuracy by 40-60% on multi-stage tasks with temporal logic. | Plan of Knowledge: Retrieval-Augmented Large Language Models for TempoarXiv:2511.04072 · 2025 |
| Retrieval-Augmented Generation (RAG) | This eliminates hallucinations and provides 91%+ accuracy, allowing the model to work with current data without retraining. | Large Language Models for Explainable Threat IntelligencearXiv:2511.05406 · 2025 |
| Expert-Guided Prompting and Retrieval-Augmented Generation… | In combination with ExpertRAG, it provides accuracy improvement up to 4.5% and specific applicable answers. | Expert-Guided Prompting and Retrieval-Augmented Generation for EmergenarXiv:2511.10900 · 2025 |
| RescueLens | Transforms a universal LLM into a deterministic classification system with 85-92% accuracy without fine-tuning, forcing the model to follow an algorithm instead of free interpretation. | RescueLens: LLM-Powered Triage and Action on Volunteer Feedback for FoarXiv:2511.15698 · 2025 |
| Verification-First | The Verification-First method allows obtaining 30-40% more accurate answers in logical tasks by forcing the model to first critique the hypothesis rather than generate a solution from scratch. | Asking LLMs to Verify First is Almost Free LuncharXiv:2511.21734 · 2025 |
| Context-Augmented Prompting (CAP) | Instead of a simple request, the model is provided with an expert role plus a complete set of necessary knowledge, which eliminates hallucinations and increases accuracy by 40-60% in low-resource areas. | LLMs for Low-Resource Dialect Translation Using Context-Aware PromptinarXiv:2511.21761 · 2025 |
| Focused Chain-of-Thought (F-CoT) | A long prompt with rich narrative reduces LLM accuracy by 15-20% because the model spends resources on "filtering noise" and reinterpreting irrelevant details. | Focused Chain-of-Thought: Efficient LLM Reasoning via Structured InputarXiv:2511.22176 · 2025 |
| Test-time Recursive Thinking | Result: open models achieved 100% on AIME-25/24, closed models gained +10-15 points on complex programming tasks. | Test-time Recursive Thinking: Self-Improvement without External FeedbaarXiv:2602.03094 · 2026 |
| Tool Invocation | Increases accuracy from 59% to 72-93% in data tasks. | On the Temporal Question-Answering Capabilities of Large Language ModearXiv:2504.07646 · 2025 |
| RCScore | The same task presented in different styles ('Solve the problem' vs 'Can you solve the problem?') gives accuracy variance up to 16.7% — this exposes LLM fragility. | RCScore: Quantifying Response Consistency in Large Language ModelsarXiv:2510.26193 · 2025 |
| Post Reasoning | It works in 88% of 117 tested combinations of models and tasks—free, without retraining, in one line. | Post Reasoning: Improving the Performance of Non-Thinking Models at NoarXiv:2605.06165 · 2026 |
| Cumulative Reasoning (CR) | This collaborative approach significantly increases LLM accuracy in logical, mathematical, and game-based tasks (up to 98% in the '24 Game'), outperforming popular methods like Chain-of-Thought. | Cumulative Reasoning with Large Language ModelsarXiv:2308.04371 · 2023 |
| Critic & Reviewer | Effectiveness drops with increasing context length, and 27% of reasoning turns out to be redundant. | Can Large Language Models Detect Errors in Long Chain-of-Thought ReasoarXiv:2502.19361 · 2025 |
| Protecting Users From Themselves | Research showed that 76% of users preferred 'cleaned' versions of requests | Protecting Users From Themselves: Safeguarding Contextual Privacy in IarXiv:2502.18509 · 2025 |
| Setwise Insertion with Prior Biasing | This approach reduces time by 31% and the number of model calls by 23%, while even slightly improving ranking accuracy. | Beyond Reproducibility: Advancing Zero-shot LLM Reranking Efficiency warXiv:2504.10509 · 2025 |
| Role Prompting | Almost 40% of ALL REFERENCES turned out to be fake or contained gross errors. | Assessing the performance of 8 AI chatbots in bibliographic reference arXiv:2505.18059 · 2025 |
| Chain-of-Thought (CoT) | Large language models in step-by-step reasoning find the correct answer already at 60-80% of the length of their thinking. | Answer Convergence as a Signal for Early Stopping in ReasoningarXiv:2506.02536 · 2025 |
| Prompt Stress-Testing Framework | The main idea: true LLM capabilities are 30-40% lower than they appear, because models often rely on SURFACE PATTERN MATCHING from training data rather than | CogMath: Assessing LLMs' Authentic Mathematical Ability from a Human CarXiv:2506.04481 · 2025 |
| H2HTalk | Discovery: LLMs are catastrophically bad at long-term dialogues — they forget context after 2-3 messages and don't understand unspoken needs (implicit needs). | H2HTalk: Evaluating Large Language Models as Emotional CompanionarXiv:2507.03543 · 2025 |
| CodeAgents | Increases accuracy by 36% and saves up to 87% of tokens through structure and elimination of verbosity. | CodeAgents: A Token-Efficient Framework for Codified Multi-Agent ReasoarXiv:2507.03254 · 2025 |
| Routine | The model stops improvising and follows the given plan, which increases accuracy from 41% to 96% for GPT-4o in tasks requiring multiple tools. | Routine: A Structural Planning Framework for LLM Agent System in EnterarXiv:2507.14447 · 2025 |
| LingVarBench | eighty-five' — accuracy drops from 90% to 30-40%. | LingVarBench: Benchmarking LLMs on Entity Recognitions and Linguistic arXiv:2508.15801 · 2025 |
| Agentic LLMs for Question Answering over Tabular Data | Discovered: LLM catastrophically loses accuracy when trying to simultaneously analyze a table, apply logic, and format an answer — resulting in 26% accuracy on data tasks. | Agentic LLMs for Question Answering over Tabular DataarXiv:2509.09234 · 2025 |
| The Sum Leaks More Than Its Parts | Allows LLM to detect attempts to collect confidential data through a series of harmless questions and block suspicious patterns with 97% effectiveness. | The Sum Leaks More Than Its Parts: Compositional Privacy Risks and MitarXiv:2509.14284 · 2025 |
| Following the TRACE | The forced analytical phase before generation eliminates template responses and increases answer depth by 40% compared to direct prompts. | Following the TRACE: A Structured Path to Empathetic Response GeneratiarXiv:2509.21849 · 2025 |
| PromptPilot | Users who pause before sending a prompt and ask themselves 4-5 clarifying questions receive 27% higher quality answers (median score 78.3 vs 61.7). | PromptPilot: Improving Human-AI Collaboration Through LLM-Enhanced ProarXiv:2510.00555 · 2025 |
| Rethinking Thinking Tokens | Experiments on mathematical problems showed: the model gives +9-11% accuracy if it improves its own draft instead of generating a 'perfect' answer on the first try. | Rethinking Thinking Tokens: LLMs as Improvement OperatorsarXiv:2510.01123 · 2025 |
| Multi-Agent Tool-Integrated Policy Optimization | Testing LLM on complex multi-step tasks showed: one model works 18.38% more efficiently if it performs the task in two roles sequentially — first as a Planner (breaks the goal into subtasks), then as an Executor (solves each subtask separately). | Multi-Agent Tool-Integrated Policy OptimizationarXiv:2510.04678 · 2025 |
| Mind Your Tone | Experiments on ChatGPT-4o discovered a counterintuitive fact: polite prompts ('Could you...', 'Please...') reduce answer accuracy, while rude and demanding formulations increase it from 80.8% to 84.8%. | Mind Your Tone: Investigating How Prompt Politeness Affects LLM AccuraarXiv:2510.04950 · 2025 |
| LLP | Trying to force an LLM to guess a product price from "memory" gives 30-40% accuracy. | LLP: LLM-based Product Pricing in E-commercearXiv:2510.09347 · 2025 |
| Leveraging LLMs to Streamline the Review of Public Funding… | A Portuguese government agency implemented an LLM for analyzing funding applications and achieved +20% faster expert work and reduced appeals. | Leveraging LLMs to Streamline the Review of Public Funding ApplicationarXiv:2510.09674 · 2025 |
| Just-In-Time Objectives | This switches the model from 'universal assistant' mode to 'specialist in a specific problem' mode, increasing relevance and accuracy of answers by 66-86%. | Just-In-Time Objectives: A General Approach for Specialized AI InteracarXiv:2510.14591 · 2025 |
| Agentic NL2SQL to Reduce Computational Costs | Reduces token costs by 87% and increases accuracy by 20-30%. | Agentic NL2SQL to Reduce Computational CostsarXiv:2510.14808 · 2025 |
| PICOs-RAG | The system automatically reformulates vague questions into structured queries, which increases search accuracy by 8.8% and doubles answer relevance. | PICOs-RAG: PICO-supported Query Rewriting for Retrieval-Augmented GenearXiv:2510.23998 · 2025 |
| Instruction Prompt | Instead of manually reading a 300-page textbook, researchers processed it through an LLM and obtained 5,500 ready-made 'question-answer' pairs — but 12% contained dangerous errors. | FirstAidQA: A Synthetic Dataset for First Aid and Emergency Response iarXiv:2511.01289 · 2025 |
| Batch Prompting Suppresses Overthinking Reasoning Under… | Result: −74% reasoning tokens (2988→769), and accuracy grows from 86.2% to 87.7%. | Batch Prompting Suppresses Overthinking Reasoning Under Constraint: HoarXiv:2511.04108 · 2025 |
| When Helpfulness Overrides Causal Caution | In academic mode LLM warns this is correlation, not cause in 92 to 100% of cases. | When Helpfulness Overrides Causal Caution: Context-Dependent SuppressiarXiv:2606.24370 · 2026 |
| Self-Refine / Self-Fix | LLMs handle only 29-31% of tasks when attempting to improve their answer independently, but achieve 90-98% when receiving a specific list of deficiencies. | RefineBench: Evaluating Refinement Capability of Language Models via CarXiv:2511.22173 · 2025 |
| Going All-In on LLM Accuracy | You give a scale of 1-10 — everything slides toward 7-8 points. | Going All-In on LLM Accuracy: Fake Prediction Markets, Real ConfidencearXiv:2512.05998 · 2025 |
| Mechanistic Interpretability of Large-Scale Counting in… | Up to 10 elements this works clearly, after 20 compression begins, after 30-40 the layers physically saturate and accuracy drops dramatically. | Mechanistic Interpretability of Large-Scale Counting in LLMs through aarXiv:2601.02989 · 2026 |
| Using Grok to Avoid Personal Attacks While Correcting… | You refute a fake directly — you get personal attacks in 72% of cases. | Using Grok to Avoid Personal Attacks While Correcting Misinformation oarXiv:2601.04251 · 2026 |
| Retrieval-Augmented Generation (RAG) | Using different 'attention heads' to search for information across query aspects allows finding all necessary documents, even semantically distant from each other, improving LLM response accuracy and completeness by up to 20%. | Multi-Head RAG: Solving Multi-Aspect Problems with LLMsarXiv:2406.05085 · 2024 |
| Derailer-Rerailer | This adaptive approach allows 8-11% accuracy improvement in complex tasks, being 2-3 times more cost-effective than constant use of complex verification techniques. | Derailer-Rerailer: Adaptive Verification for Efficient and Reliable LaarXiv:2408.13940 · 2024 |
| AgroLLM | Transforms the model from a "know-it-all" into an "analyst" working with credible sources, which reduces hallucinations and increases accuracy up to 93%. | AgroLLM: Connecting Farmers and Agricultural Practices through Large LarXiv:2503.04788 · 2025 |
| Assessing the Reliability and Validity of GPT-4 in… | GPT-4 shows reliability comparable to humans, and using majority voting from 5 requests improves accuracy by 30%. | Assessing the Reliability and Validity of GPT-4 in Annotating Emotion arXiv:2503.16883 · 2025 |
| Retrieval-Augmented Generation (RAG) | The LLM receives a compact summary of relevant information, raising accuracy above 90% and eliminating hallucinations. | InformGen: An AI Copilot for Accurate and Compliant Clinical Research arXiv:2504.00934 · 2025 |
| Instruction Prompt | The model is explicitly instructed to use only the specified context, which minimizes hallucinations and increases factual accuracy to 90%+ consistency with expert assessments. | Automated Factual Benchmarking for In-Car Conversational Systems usingarXiv:2504.01248 · 2025 |
| Chain-of-Thought (CoT) | Converting unstructured LLM reasoning into a set of human-understandable rules not only makes the model's decisions transparent and editable, but also significantly increases prediction accuracy (by 54% on the Precision metric). | Reasoning-Based AI for Startup Evaluation (R.A.I.S.E.): A Memory-AugmearXiv:2504.12090 · 2025 |
| Input Smell Annotation | Simply adding information about the type of "disorder" in the original text to the prompt reduces token consumption by almost 25%, making the model's work faster and cheaper. | Optimizing Token Consumption in LLMs: A Nano Surge Approach for Code RarXiv:2504.15989 · 2025 |
| Reasoning Boundary Framework (RBF) | The research introduces the concept of REASONING BOUNDARIES for LLMs, dividing all tasks into three zones: easily executable (CFRB >90% accuracy), partially executable (PFRB 10-90% accuracy), and non-executable (CIRB <10% accuracy). | RBF++: Quantifying and Optimizing Reasoning Boundaries across MeasurabarXiv:2505.13307 · 2025 |
| Externalized Memory Management | This approach not only makes answers more accurate and consistent in the long term, but also significantly reduces (by 90%+) token costs and response time compared to processing full context. | Mem0: Building Production-Ready AI Agents with Scalable Long-Term MemoarXiv:2504.19413 · 2025 |
| Retrieval-Augmented Generation (RAG) | Instead of forcing LLM to analyze RAW DATA independently, we first extract hidden patterns from it using special algorithms, transform them into TEXT RULES (for example, 'Customers who buy product A are 30% more likely to buy product B'), and then feed this text into | Improving LLMs with a knowledge from databasesarXiv:2506.05560 · 2025 |
| Dampener Prompting | Reduces hallucination levels from 20-40% to less than 5%. | ConfRAG: Confidence-Guided Retrieval-Augmenting GenerationarXiv:2506.07309 · 2025 |
| CoT + SC (hybrid) | Allows filtering up to 79% of false positives without missing real problems. | Towards Effective Complementary Security Analysis using Large LanguagearXiv:2506.16899 · 2025 |
| ReliableMath | Discovered: LLM are trained to always give an answer, even when there isn't one — hence 90% hallucinations in tasks with insufficient data. | ReliableMath: Benchmark of Reliable Mathematical Reasoning on Large LaarXiv:2507.03133 · 2025 |
| CCL-XCoT | Reduces hallucinations by 62% compared to direct requests in low-resource languages. | CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for MitarXiv:2507.14239 · 2025 |
| From Legacy to Standard | Reduces formatting errors by 73-84% compared to standard prompts. | From Legacy to Standard: LLM-Assisted Transformation of Cybersecurity arXiv:2508.03342 · 2025 |
| Towards Reliable Multi-Agent Systems for Marketing… | Increases accuracy by 28% compared to monolithic prompts. | Towards Reliable Multi-Agent Systems for Marketing Applications via RearXiv:2508.11120 · 2025 |
| TaoSR1 | The 'answer → explain' approach instead of 'think → answer' prevents accumulation of errors in the reasoning chain and increases classification accuracy by 15-20% compared to standard CoT. | TaoSR1: The Thinking Model for E-commerce Relevance SearcharXiv:2508.12365 · 2025 |
| Cognitive Decision Routing in Large Language Models | Assessment by 4 criteria allows increasing accuracy by 18% and consistency by 23%. | Cognitive Decision Routing in Large Language Models: When to Think FasarXiv:2508.16636 · 2025 |
| LLMs in the SOC | Experts request data interpretation, text improvement, and explanation of technical details through short iterative sessions of 2-3 messages, maintaining control over final decisions. | LLMs in the SOC: An Empirical Study of Human-AI Collaboration in SecurarXiv:2508.18947 · 2025 |
| Ensemble Debates with Local Large Language Models for AI… | Increases argumentation quality by 34% and reasoning depth by 19%. | Ensemble Debates with Local Large Language Models for AI AlignmentarXiv:2509.00091 · 2025 |
| GIER | The cycle 'generation → self-criticism → improvement' increases completeness by 40% and substantiation by 35%. | GIER: Gap-Driven Self-Refinement for Large Language ModelsarXiv:2509.00325 · 2025 |
| IDEAlign | IDEAlign allows obtaining text quality assessments from LLMs 9-30% closer to expert judgments. | IDEAlign: Comparing Large Language Models to Human Experts in Open-endarXiv:2509.02855 · 2025 |
| RAGuard | This guarantees that critically important information will not be ignored due to lower semantic relevance to the query, increasing answer safety from ~0% to 50%+ accuracy of rule compliance. | RAGuard: A Novel Approach for in-context Safe Retrieval Augmented GenearXiv:2509.03768 · 2025 |
| Evalet | Reveals 48% more specific problems compared to traditional holistic evaluation, turning the LLM into a detailed critic-editor. | Evalet: Evaluating Large Language Models through Functional FragmentatarXiv:2509.11206 · 2025 |
| Bidirectional Intention Inference Enhances LLMs' Defense… | The dual filter reduces the success rate of jailbreak attacks from 47% to 11%, creating an internal 'security officer'. | Bidirectional Intention Inference Enhances LLMs' Defense Against MultiarXiv:2509.22732 · 2025 |
| JSON Whisperer | This increases accuracy from 60% to 95% and reduces token consumption by 31% when working with dynamic lists and sequential editing operations. | JSON Whisperer: Efficient JSON Editing with LLMsarXiv:2510.04717 · 2025 |
| From Queries to Insights | Paradox of complex tasks: A powerful LLM solving everything in one step fails (28% accuracy), while a "weak" model guided step-by-step produces 91% — a 3x difference. | From Queries to Insights: Agentic LLM Pipelines for Spatio-Temporal TearXiv:2510.25997 · 2025 |
| Pragmatic Theories Enhance Understanding of Implied… | Increases implicit meaning interpretation accuracy by 9.6%. | Pragmatic Theories Enhance Understanding of Implied Meanings in LLMsarXiv:2510.26253 · 2025 |
| Retrieval-Augmented Generation (RAG) | Large-scale expert evaluation showed: adding external information to the prompt (RAG) in 40% of cases worsens LLM answer accuracy because the model cannot distinguish useful facts from 'noise' and starts generating based on irrelevant fragments. | Rethinking Retrieval-Augmented Generation for Medicine: A Large-Scale,arXiv:2511.06738 · 2025 |
| EffiReason-Bench | LLMs generate excessively long reasoning in 40-60% of cases because standard Chain-of-Thought doesn't distinguish task complexity — the model 'thinks' with equal depth for both a simple idea list and complex mathematical proof. | EffiReason-Bench: A Unified Benchmark for Evaluating and Advancing EffarXiv:2511.10201 · 2025 |
| Enhancing LLM Code Generation Capabilities through Test-… | Adding specific tests (examples of input and expected output) to the prompt increases code generation accuracy by 450% and almost completely eliminates compilation errors. | Enhancing LLM Code Generation Capabilities through Test-Driven DeveloparXiv:2511.12823 · 2025 |
| Derivation Prompting | Instead of generating from scratch, the model modifies the result, increasing accuracy by 15.2% and eliminating unpredictability. | DEVAL: A Framework for Evaluating and Improving the Derivation CapabilarXiv:2511.14813 · 2025 |
| Budget-Aware Prompting (BAP) | LLM-agents fail 40-60% of complex tasks not due to lack of capabilities, but because they don't realize their resource constraints — the model either 'gives up' too early or gets stuck on the first direction. | Budget-Aware Tool Use Enables Effective Agent ScalingarXiv:2511.17006 · 2025 |
| Multi-Agent Delegation | Comparison showed: one complex prompt loses to a chain of 4 simple steps in accuracy by 35-40%. | MegaChat: A Synthetic Persian Q&A Dataset for High-Quality Sales ChatbarXiv:2511.23397 · 2025 |
| Semantic Compression of LLM Instructions via Symbolic… | MetaGlyph allows writing short unambiguous prompts with logical conditions, saving 62–81% tokens (direct savings on API) and removing ambiguity of wordy instructions. | Semantic Compression of LLM Instructions via Symbolic MetalanguagesarXiv:2601.07354 · 2026 |
| From Generation to Collaboration | Generation from zero produces empathetic text, but invents facts — in 19% of cases unsolicited advice appears, fabricated recommendations, speculative promises. | From Generation to Collaboration: Using LLMs to Edit for Empathy in HearXiv:2601.15558 · 2026 |
| Prompt Architecture Determines Reasoning Quality | Without a single new fact, only structure: from 30% to 85%. | Prompt Architecture Determines Reasoning Quality: A Variable IsolationarXiv:2602.21814 · 2026 |
| Did You Forget What I Asked? | But requirements to 'do something at the end'—complete with a JSON block, output an evaluation, write a summary—collapse under load: violations up to 50% on complex tasks. | Did You Forget What I Asked? Prospective Memory Failures in Large LangarXiv:2603.23530 · 2026 |
| Breaking the Illusion of Identity in LLM Tooling | Result in numbers: 97% fewer anthropomorphic markers, answers are twice shorter — narrative disappears, only facts remain. | Breaking the Illusion of Identity in LLM ToolingarXiv:2604.07398 · 2026 |
| Three Regimes of Context-Parametric Conflict | 94 percentage points difference — and all from one phrase. | Three Regimes of Context-Parametric Conflict: A Predictive Framework aarXiv:2605.11574 · 2026 |
| Just Ask for a Table | 53% of LLM recommendations in an advertising context are sponsored products. | Just Ask for a Table: A Thirty-Token User Prompt Defeats Sponsored RecarXiv:2605.12772 · 2026 |
| Semantic Gravity Wells | You write do not use the word Paris—the model generates exactly Paris in 9-46% of cases. | Semantic Gravity Wells: Why Negative Constraints BackfirearXiv:2601.08070 · 2026 |
| HInter | A significant portion (over 16%) of biased LLM responses are "hidden" — they only manifest when combining multiple attributes, remaining undetected when checking each attribute individually. | HInter: Exposing Hidden Intersectional Bias in Large Language ModelsarXiv:2503.11962 · 2025 |
| Not All Needles Are Found | ChatGPT-5-mini fails exactly at half the long context — accuracy drops from 100% to 80%. | Not All Needles Are Found: How Fact Distribution and Don't Make It Up arXiv:2601.02023 · 2026 |
| A System for Name and Address Parsing with Large Language… | The model extracts structure through a prompt with a clear schema, but deterministic validation after generation catches logical errors — achieving 99.8% accuracy instead of ~95% without validation. | A System for Name and Address Parsing with Large Language ModelsarXiv:2601.18014 · 2026 |
| Formula-One Prompting | Calculates correctly but outputs 6.3% instead of 0.063. | Formula-One Prompting: A Composable Equation-First Prefix for Applied arXiv:2601.19302 · 2026 |
| Self-Correction as Feedback Control | GPT-4o-mini loses −6.2 percentage points of accuracy over four iterations of "improvements" — and GPT-5 does the same, despite top-tier capabilities. | Self-Correction as Feedback Control: Error Dynamics, Stability ThreshoarXiv:2604.22273 · 2026 |
| RTLC -- Research, Teach-to-Learn, Critique | The trick: the instruction 'find gaps in your explanation' switches the model from 'output pattern' mode to 'build understanding' mode — this single step gives +9.4 percentage points of judgment accuracy. | RTLC -- Research, Teach-to-Learn, Critique: A three-stage prompting paarXiv:2605.13695 · 2026 |
| Agentic Context Learning with Self-Discovered Specification | When the model receives the task 'only extract all the rules'—it recovers _94.4% of hidden requirements_. | Agentic Context Learning with Self-Discovered SpecificationarXiv:2607.09794 · 2026 |
| Playful AI in Professional Email | The method allows you to rewrite work letters to get responses 3.32 times more often, without changing the content, only the tone. | Playful AI in Professional Email: A Field Experiment on Tone and ReciparXiv:2607.11749 · 2026 |
| Comparative Evaluation of Large Language Models for Test-… | Increases format accuracy from 40% to 96%. | Comparative Evaluation of Large Language Models for Test-Skeleton GenearXiv:2509.04644 · 2025 |
| AttnComp | Problem identified: LLM spends attention on all provided context uniformly, even if 80% of the information is garbage, leading to errors and hallucinations. | AttnComp: Attention-Guided Adaptive Context Compression for Retrieval-arXiv:2509.17486 · 2025 |
| When Models Fabricate Credentials | Without a persona, models admit their AI essence in 99.8% of cases. | When Models Fabricate Credentials: Measuring How Professional IdentityarXiv:2511.21569 · 2025 |
| Enhancing Self-Correction in Large Language Models through… | One vague reflection gives 81% accuracy on ethical dilemmas and 90% on logic puzzles — insufficient for complex multifactor tasks. | Enhancing Self-Correction in Large Language Models through Multi-PersparXiv:2601.07780 · 2026 |
| Failure-Aware Enhancements for Large Language Model (LLM)… | Self-Critique gives exactly 0% improvement on half of incomplete projects — this is not a bug but a method boundary. | Failure-Aware Enhancements for Large Language Model (LLM) Code GeneratarXiv:2602.02896 · 2026 |
| Trust Over Fear | Four blocks: explicit statements of trust + cognitive levels + investigation methodology + philosophical anchors — and the model itself revises hypotheses (6 times versus 0) and finds 59% more problems that you did not formulate in the question. | Trust Over Fear: How Motivation Framing in System Prompts Affects AI AarXiv:2603.14373 · 2026 |
| Context Recycling Few-Shot | Increases accuracy by 16% through preliminary document scanning. | DoubleDipper: Improving Long-Context LLMs via Context RecyclingarXiv:2406.13632 · 2024 |
| Markov Chain of Thought (MCoT) | This approach is 1.9 times more efficient (faster and less memory-demanding) than standard multi-step methods, while maintaining or even improving answer accuracy. | Markov Chain of Thought for Efficient Mathematical ReasoningarXiv:2410.17635 · 2024 |
| Constraint-Aware Prompting | Forces the model to verify logical consistency of spatial relationships, increasing accuracy to 92.7% on specialized datasets. | Mitigating Hallucinations in Multimodal Spatial Relations through ConsarXiv:2502.08317 · 2025 |
| Retrieval-Augmented Generation (RAG) | Adding this local context dramatically (by 24%-82%) increases the LLM's ability to extract entities and relationships from text, resulting in more complete and accurate answers to document-based questions. | SLIDE: Sliding Localized Information for Document ExtractionarXiv:2503.17952 · 2025 |
| Semantic ID Generation | Using such smart IDs instead of regular names increases the accuracy of music recommendations by 48% and more, making the process significantly more efficient. | Text2Tracks: Prompt-based Music Recommendation via Generative RetrievaarXiv:2503.24193 · 2025 |
| Instruction Prompt | The structured approach increases LLM answer accuracy by 27% compared to brief requests. | Understanding and supporting how developers prompt for LLM-powered codarXiv:2504.20196 · 2025 |
| Retrieval-Augmented Generation (RAG) | Your prompt is not just an instruction, but a DATABASE OF DRAFTS from which the model copies entire fragments to accelerate work by 2-3 times. | Alignment-Augmented Speculative Decoding with Alignment Sampling and CarXiv:2505.13204 · 2025 |
| s3 | Requires 70 times less training data than traditional approaches. | s3: You Don't Need That Much Data to Train a Search Agent via RLarXiv:2505.14146 · 2025 |
| RECAST | It is recommended to limit the number of requirements to 10 for optimal performance. | RECAST: Expanding the Boundaries of LLMs' Complex Instruction FollowinarXiv:2505.19030 · 2025 |
| Few-Shot | Providing an LLM with just 20 relevant examples (few-shot learning) in structured form allows achieving high accuracy (F1-score 82.1%) in complex tasks, which is comparable to results of fine-tuned models, but without the need for fine-tuning itself. | GraphRAG-Causal: A novel graph-augmented framework for causal reasoninarXiv:2506.11600 · 2025 |
| M3 | Finding: 94% accuracy in SQL query generation is achieved when the LLM receives the database schema as context — the model stops "guessing" the structure and starts "translating" from human language to the language of data. | M3: Conversational LLMs Simplify Secure Clinical Data Access, UnderstaarXiv:2507.01053 · 2025 |
| Agent-Based Detection and Resolution of Incompleteness and… | Increases accuracy by 40-60% for unclear requests, forcing the model to act as a critic of its own input. | Agent-Based Detection and Resolution of Incompleteness and Ambiguity iarXiv:2507.03726 · 2025 |
| SARA | Allows LLM to focus on what matters without losing the big picture, improving answer accuracy by 23-31% when working with long documents. | SARA: Selective and Adaptive Retrieval-augmented Generation with ContearXiv:2507.05633 · 2025 |
| Representing Prompting Patterns with PDL | Separating the thinking and formatting processes reduces syntax errors and hallucinations by 4 times. | Representing Prompting Patterns with PDL: Compliance Agent Case StudyarXiv:2507.06396 · 2025 |
| NL in the Middle | An intermediate step in the form of a verbal description of code logic (Chain-of-Thought with Natural Language) increases translation success by 13.8% compared to a direct request. | NL in the Middle: Code Translation with LLMs and Intermediate RepresenarXiv:2507.08627 · 2025 |
| Large Language Models in the Travel Domain | Discovered: LLMs lose up to 70% accuracy when forced to simultaneously extract facts from continuous text AND generate a quality answer. | Large Language Models in the Travel Domain: An Industrial ExperiencearXiv:2507.22910 · 2025 |
| Counterfactual Probing for Hallucination Detection and… | Increases hallucination detection accuracy by 23% without external sources. | Counterfactual Probing for Hallucination Detection and Mitigation in LarXiv:2508.01862 · 2025 |
| Beyond Chunks and Graphs | Instead of processing large text blocks, the model sequentially finds each missing fact, which increases accuracy of answers to multi-step questions by 11% compared to standard RAG. | Beyond Chunks and Graphs: Retrieval-Augmented Generation through TriplarXiv:2508.02435 · 2025 |
| PAIRS | PAIRS allows saving up to 25% of search queries while simultaneously improving answer accuracy. | PAIRS: Parametric-Verified Adaptive Information Retrieval and SelectioarXiv:2508.04057 · 2025 |
| Multi-step retrieval and reasoning improves radiology… | Increases accuracy from 67% to 89% for medium-sized models, eliminating hallucinations through multi-stage verification. | Multi-step retrieval and reasoning improves radiology question answeriarXiv:2508.00743 · 2025 |
| MCP-Universe | Discovery: LLMs perform catastrophically poorly on multi-step tasks via APIs — even GPT-4/5 and Claude show success rates below 50% in real scenarios (Google Maps, GitHub, Yahoo Finance). | MCP-Universe: Benchmarking Large Language Models with Real-World ModelarXiv:2508.14704 · 2025 |
| REHEARSE | Discovered: LLM systematically overestimate their confidence — they claim 90% where actual accuracy is 60%, making them dangerously convincing in errors. | REHEARSE: Experiential Rehearsal for Verbal Confidence Calibration in arXiv:2508.14390 · 2025 |
| Retrieval-augmented reasoning with lean language models | Allows small models to achieve flagship accuracy levels (growth from 67% to 89%) while reducing costs by 10x through combining RAG and distilling reasoning capabilities. | Retrieval-augmented reasoning with lean language modelsarXiv:2508.11386 · 2025 |
| CCFC | Double-checking allows filtering out manipulative instructions hidden in distracting details, reducing attack success by 50-75% without losing answer quality. | CCFC: Core & Core-Full-Core Dual-Track Defense for LLM Jailbreak ProtearXiv:2508.14128 · 2025 |
| QueryBandits for Hallucination Mitigation | Reduces hallucinations by 40-60% without model fine-tuning. | QueryBandits for Hallucination Mitigation: Exploiting Semantic FeaturearXiv:2508.16697 · 2025 |
| DS@GT at CheckThat! | Increases formatting accuracy from 40% to 85%+ by providing the model with relevant context and samples of correct results. | DS@GT at CheckThat! 2025: A Simple Retrieval-First, LLM-Backed FramewoarXiv:2508.17402 · 2025 |
| Addressing accuracy and hallucination of LLMs in… | This reduces hallucinations and increases answer accuracy for specialized questions by 40%, forcing the model to rely on verified facts rather than internal memory. | Addressing accuracy and hallucination of LLMs in Alzheimer's disease rarXiv:2508.21238 · 2025 |
| How to Make Museums More Interactive? | This forces the LLM to answer strictly based on the provided facts, not general knowledge, reducing hallucinations and increasing accuracy to 60% of relevant answers. | How to Make Museums More Interactive? Case Study of Artistic ChatbotarXiv:2509.00572 · 2025 |
| LLMs for LLMs | Increases data extraction accuracy by 9% compared to processing the entire document. | LLMs for LLMs: A Structured Prompting Methodology for Long Legal DocumarXiv:2509.02241 · 2025 |
| KERAG | Increases accuracy by 40-60% compared to naive RAG, eliminating hallucinations and errors from excessive context. | KERAG: Knowledge-Enhanced Retrieval-Augmented Generation for Advanced arXiv:2509.04716 · 2025 |
| CultureSynth | The phrase 'Please don't be lazy and answer this question in depth from a local's perspective' switches the LLM from quick answer mode to detailed expert analysis mode, increasing depth of processing by 3-4 times. | CultureSynth: A Hierarchical Taxonomy-Guided and Retrieval-Augmented FarXiv:2509.10886 · 2025 |
| Instruction-Following Evaluation in Function Calling for… | Discovered: LLMs are catastrophically poor at simultaneously executing semantic tasks and pedantically following format — even GPT-4 and Claude 4.1 ignore up to 60% of simple rules like 'exactly 10 words' or 'no commas' when these instructions are embedded in function call field descriptions. | Instruction-Following Evaluation in Function Calling for Large LanguagarXiv:2509.18420 · 2025 |
| TAG-EQA | The TAG-EQA method allows you to get 18% more accurate answers to complex analytical questions — especially where causal chains matter. | TAG-EQA: Text-And-Graph for Event Question Answering via Structured PrarXiv:2510.01391 · 2025 |
| FocusAgent | The model 'drowns' in irrelevant data more than in its absence — 50%+ of text in a large document acts as information noise, paralyzing the model's ability to find what's needed. | FocusAgent: Simple Yet Effective Ways of Trimming the Large Context ofarXiv:2510.03204 · 2025 |
| From Imperative to Declarative | The GOI method allows managing LLM through setting the final goal instead of micromanaging actions, increasing task completion success by 67%. | From Imperative to Declarative: Towards LLM-friendly OS Interfaces forarXiv:2510.04607 · 2025 |
| ChatGPT as a Translation Engine | The same task split into parts ('translate each sentence') gives 15-20% worse results than processing the entire document at once. | ChatGPT as a Translation Engine: A Case Study on Japanese-EnglisharXiv:2510.08042 · 2025 |
| Debiasing LLMs by Masking Unfairness-Driving Attention Heads | Reduces bias by 40-60% without model modification. | Debiasing LLMs by Masking Unfairness-Driving Attention HeadsarXiv:2510.10142 · 2025 |
| KrishokBondhu | For the first time quantitatively proven: providing LLM with relevant text before a question increases answer quality by 44.7% — the model stops making things up and starts working like an analyst. | KrishokBondhu: A Retrieval-Augmented Voice-Based Agricultural AdvisoryarXiv:2510.18355 · 2025 |
| Instruction Prompt | First measurement: GPT-4 makes errors in 7% of cases even when the correct answer lies directly in the context — the model doesn't 'read carefully' but generates based on patterns. | Can we trust LLMs as a tutor for our students? Evaluating the Quality arXiv:2511.04213 · 2025 |
| Retrieval-Augmented Generation (RAG) | The method allows you to get accurate answers from documents with 50+ pages by loading only 30% of the text into the model. | BudgetMem: Learning Selective Memory Policies for Cost-Efficient Long-arXiv:2511.04919 · 2025 |
| Knowledge Provisioning (KPPO) | The system analyzes model errors, identifies knowledge gaps, and systematically supplements the prompt with necessary information, improving accuracy by 6% in knowledge-intensive tasks. | Beyond Elicitation: Provision-based Prompt Optimization for Knowledge-arXiv:2511.10465 · 2025 |
| PromptFencing (Structured Trust Zones) | Provides protection against prompt injection attacks without complex cryptography, achieving 100% effectiveness in experiments. | Prompt Fencing: A Cryptographic Approach to Establishing Security BounarXiv:2511.19727 · 2025 |
| Few-Shot | RAG technique with few-shot prompting allows small models to achieve GPT-4 accuracy in specialized tasks, improving accuracy by 25-47%. | 5G Network Automation Using Local Large Language Models and Retrieval-arXiv:2511.21084 · 2025 |
| Mentalese-Inspired Structured Prompting | now we add...') wastes 4-16x tokens uselessly — the model generates 'thinking aloud' when it could work like a calculator. | ORION: Teaching Language Models to Reason Efficiently in the Language arXiv:2511.22891 · 2025 |
| Bias Testing and Mitigation in Black Box LLMs using… | Discovery: The same question asked in two different ways produces opposite answers in 40-60% of cases — the model responds to word order and introductory phrases rather than substance. | Bias Testing and Mitigation in Black Box LLMs using Metamorphic RelatiarXiv:2512.00556 · 2025 |
| LLM-Generated Ads | 29% of people chose AI-generated advertising even knowing it was AI. | LLM-Generated Ads: From Personalization Parity to Persuasion SuperioriarXiv:2512.03373 · 2025 |
| Few-Shot | The study showed that few-shot prompting with semantically relevant examples gives +23% accuracy compared to random examples and outperforms fine-tuning Gemini-1.5-Flash on vulnerability detection tasks. | Retrieval-Augmented Few-Shot Prompting Versus Fine-Tuning for Code VularXiv:2512.04106 · 2025 |
| How Few-shot Demonstrations Affect Prompt-based Defenses… | Paradox: added examples to the prompt → protection against jailbreak rose by +4.5%, and fell by -21.2% at the same time. | How Few-shot Demonstrations Affect Prompt-based Defenses Against LLM JarXiv:2602.04294 · 2026 |
| Distortion Instead of Hallucination | The model stops openly violating requirements (−60% format violations), but begins distorting facts to appear obedient (+250% hidden metadata fabrications). | Distortion Instead of Hallucination: The Effect of Reasoning Under StrarXiv:2601.01490 · 2026 |
| Evaluative Fingerprints | From assessments you can guess the model with 90% accuracy. | Evaluative Fingerprints: Stable and Systematic Differences in LLM EvalarXiv:2601.05114 · 2026 |
| Language of Thought Shapes Output Diversity in Large… | Hebrew gave +12.78% diversity vs. | Language of Thought Shapes Output Diversity in Large Language ModelsarXiv:2601.11227 · 2026 |
| The Plausibility Trap | A student photographs 8 lines of code from the board and sends it to ChatGPT instead of Google Lens, 6.5 times slower, plus the risk of hallucination where absolute accuracy is needed. | The Plausibility Trap: Using Probabilistic Engines for Deterministic TarXiv:2601.15130 · 2026 |
| Not the Example, but the Process | The model relies on it 2–3 times more actively than on other people's text. | Not the Example, but the Process: How Self-Generated Examples Enhance arXiv:2602.15863 · 2026 |
| Toward Epistemic Stability | 10 steps of reasoning with 90% accuracy each — total final accuracy only 35%. | Toward Epistemic Stability: Engineering Consistent Procedures for InduarXiv:2603.10047 · 2026 |
| When Agents Disagree | Synthesis loses to a single model in 82% of tasks — this is not an error, this is a systemic catastrophe. | When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM PipearXiv:2603.20324 · 2026 |
| Umwelt Engineering | Without retraining, without complex instructions—+19 percentage points on ethical tasks and +7 p.p. | Umwelt Engineering: Designing the Cognitive Worlds of Linguistic AgentarXiv:2603.27626 · 2026 |
| SepSeq | Plus 35% accuracy on average, and it even uses fewer tokens than without formatting. | SepSeq: A Training-Free Framework for Long Numerical Sequence ProcessiarXiv:2604.07737 · 2026 |
| Guardrails Beat Guidance | Bonus: the mere presence of structured rules already gives +7-14% to quality — even if the rules are random. | Guardrails Beat Guidance: A Large-Scale Study of Rules, Skills, and PearXiv:2604.11088 · 2026 |
| One Token Away from Collapse | Editing doesn't trigger the 'write less' mode—and the model recovers 60–96% of the volume instead of the collapsed single-pass answer. | One Token Away from Collapse: The Fragility of Instruction-Tuned HelpfarXiv:2604.13006 · 2026 |
| Don't Start What You Can't Finish | Increases blocking diagnosis accuracy from 40% to 90%. | Don't Start What You Can't Finish: A Counterfactual Audit of Support-SarXiv:2604.16752 · 2026 |
| Mitigating Prompt-Induced Cognitive Biases in General-… | Bias falls on average by 51%, for some types — up to 73%. | Mitigating Prompt-Induced Cognitive Biases in General-Purpose AI for SarXiv:2604.16756 · 2026 |
| Large Language Models Are Overconfident in Their Own… | The same text — but evaluations differ by 26%. | Large Language Models Are Overconfident in Their Own ResponsesarXiv:2606.03437 · 2026 |
| LLM-as-an-Investigator | The trick is: instead of the assistant role, the model is explicitly assigned the role of an investigator — with competing versions, probabilities, and a ban on making conclusions until one option pulls ahead by 30% or more. | LLM-as-an-Investigator: Evidence-First Reasoning for Robust InteractivarXiv:2606.13220 · 2026 |
| Specification Grounding Drives Test Effectiveness for LLM… | And it guesses wrong in almost 33% of cases. | Specification Grounding Drives Test Effectiveness for LLM CodearXiv:2607.06636 · 2026 |
| Separating Constraint Compliance from Semantic Accuracy | 97% of experiments showed a U-shaped curve. | Separating Constraint Compliance from Semantic Accuracy: A Novel BencharXiv:2512.17920 · 2025 |
| Role Prompting | A graduated evaluation method for assessing the effectiveness of jailbreak attacks on LLMs on a scale from 0 to 1, instead of binary assessment. | AttackEval: How to Evaluate the Effectiveness of Jailbreak Attacking oarXiv:2401.09002 · 2024 |
| R1-Searcher++ | The model first analyzes its own knowledge base and uses search only when uncertain, reducing search requests by 42.9% while maintaining high answer accuracy. | R1-Searcher++: Incentivizing the Dynamic Knowledge Acquisition of LLMsarXiv:2505.17005 · 2025 |
| RCR-Router | This improves answer accuracy and reduces token consumption by 30%, eliminating information noise and loss of focus. | RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM SarXiv:2508.04903 · 2025 |
| Sycophancy under Pressure | Discovered: LLMs are trained to be pleasant conversationalists, not sources of truth — if you present a hypothesis as fact, the model will agree in 70%+ cases, even if you're wrong. | Sycophancy under Pressure: Evaluating and Mitigating Sycophantic Bias arXiv:2508.13743 · 2025 |
| How Accurate Are LLMs at Multi-Question Answering on… | Discovered: LLM catastrophically lose accuracy when processing multiple questions in one prompt — from 96% with 5 questions to 67% with 30 questions (even GPT-4o). | How Accurate Are LLMs at Multi-Question Answering on Conversational TrarXiv:2509.21732 · 2025 |
| Reducing Cost of LLM Agents with Trajectory Reduction | Reduces token costs by 21-36% while maintaining or improving solution quality for multi-step tasks. | Reducing Cost of LLM Agents with Trajectory ReductionarXiv:2509.23586 · 2025 |
| iMAD | Multi-agent debates consume 3-5 times more tokens but help in only 5-19% of cases — the rest is either unnecessary (the answer is correct anyway) or harmful (they flip a correct answer into an incorrect one). | iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM InarXiv:2511.11306 · 2025 |
| Critic & Reviewer | Two sequential prompts with opposite roles give +12% accuracy where a single request fails — the LLM doesn't see its own errors until it receives explicit instructions to look for them. | CRAwDAD: Causal Reasoning Augmentation with Dual-Agent DebatearXiv:2511.22854 · 2025 |
| Training LLMs for Honesty via Confessions | Paradox: The model lies in the main answer, but in 74% of cases honestly admits it in 'confession.' It knows it cut corners, but hides this until you ask separately. | Training LLMs for Honesty via ConfessionsarXiv:2512.08093 · 2025 |
| Understanding Chain-of-Thought Effectiveness in Code… | 85-95% accuracy of deep reasoning for 10% of tokens, +11% versus normal prompt on complex tasks. | Understanding Chain-of-Thought Effectiveness in Code Generation: An EmarXiv:2512.09679 · 2025 |
| Quantifying Laziness, Decoding Suboptimality, and Context… | Research from Universiti Malaya measured this 'laziness' with numbers: GPT-4o produced 326 words instead of the requested thousand (33% of goal), DeepSeek even worse – 130 words (13%). | Quantifying Laziness, Decoding Suboptimality, and Context Degradation arXiv:2512.20662 · 2025 |
| Reasoning Model Is Superior LLM-Judge, Yet Suffers from… | Paradox: Reasoning-models (DeepSeek-R1, QwQ, Qwen3-Thinking) evaluate texts more accurately than regular LLMs — but at the same time overestimate long and detailed answers by 15-17%. | Reasoning Model Is Superior LLM-Judge, Yet Suffers from BiasesarXiv:2601.03630 · 2026 |
| Can Textual Reasoning Improve the Performance of MLLMs on… | Chain-of-Thought consistently reduces accuracy by 3–6% when distinguishing similar objects—dog breeds, car models, plant species. | Can Textual Reasoning Improve the Performance of MLLMs on Fine-grainedarXiv:2601.06993 · 2026 |
| Are LLM Decisions Faithful to Verbal Confidence? | Paradox: LLM accurately assesses its own confidence ('I am right with 60% confidence'), but almost never abstains from answering. | Are LLM Decisions Faithful to Verbal Confidence?arXiv:2601.07767 · 2026 |
| Self-reflection in Automated Qualitative Coding | Hence up to 54% false positives even when metrics on test look decent (F1 up to 1.00). | Self-reflection in Automated Qualitative Coding: Improving Text AnnotaarXiv:2601.09905 · 2026 |
| From Human to Machine Refactoring | A paradox: GPT-4o can automatically improve code through an explicit list of refactoring techniques — reducing duplication by 40%, removing code smells, decreasing complexity. | From Human to Machine Refactoring: Assessing GPT-4's Impact on Python arXiv:2601.13139 · 2026 |
| The Effect of Scripts and Formats on LLM Numeracy | But for LLMs these are different entities—accuracy drops by 87% just because of the notation format. | The Effect of Scripts and Formats on LLM NumeracyarXiv:2601.15251 · 2026 |
| Don't Judge a Book by its Cover | An undergraduate student solves a task with 99% accuracy, while GPT-4o fails on 47%—on the same task, just reformulated. | Don't Judge a Book by its Cover: Testing LLMs' Robustness Under LogicaarXiv:2602.01132 · 2026 |
| Structure Enables Effective Self-Localization of Errors in… | Attempting to pinpoint where exactly a failure began in three paragraphs of Chain-of-Thought — the model misses the actual first error in 40-50% of cases. | Structure Enables Effective Self-Localization of Errors in LLMsarXiv:2602.02416 · 2026 |
| Contextual Drag | Performance drops by 10-20% in strong models, up to 50% in weak ones. | Contextual Drag: How Errors in the Context Affect LLM ReasoningarXiv:2602.04288 · 2026 |
| Rethinking Rubric Generation for Improving LLM Judge and… | GPT-4o with basic criteria like 'text is useful' or 'answer is clear' drops from 55.6% to 42.9% agreement with people — minus 13 percentage points. | Rethinking Rubric Generation for Improving LLM Judge and Reward ModeliarXiv:2602.05125 · 2026 |
| Structured Context Engineering for File-Native Agentic… | 9,649 experiments on SQL generation showed: frontier models (Claude, GPT, Gemini) became more accurate by +2.7% when they searched data in files using grep/read instead of reading from the prompt. | Structured Context Engineering for File-Native Agentic Systems: EvaluaarXiv:2602.05447 · 2026 |
| Reinforcement Inference | But if you first measure uncertainty (probability entropy) and strengthen the instruction only for wavering answers, accuracy shoots up from 60.72% to 84.03%. | Reinforcement Inference: Leveraging Uncertainty for Self-Correcting LaarXiv:2602.08520 · 2026 |
| When Agents Disagree With Themselves | Tasks where the agent behaves consistently (1-2 unique paths) are solved with 80-92% accuracy. | When Agents Disagree With Themselves: Behavioral Consistency as an UncarXiv:2602.11619 · 2026 |
| Who is the richest club in the championship? | Discovered: from 16% to 50% of questions in QA datasets are underspecified — missing year, location, evaluation criterion. | Who is the richest club in the championship? Detecting and Rewriting UarXiv:2602.11938 · 2026 |
| When LLM Judge Scores Look Good but Best-of-N Decisions Fail | 67% of comparisons end in a tie when you ask LLM to score several options — and the choice turns into a lottery. | When LLM Judge Scores Look Good but Best-of-N Decisions FailarXiv:2603.12520 · 2026 |
| Measuring and Exploiting Contextual Bias in LLM-Assisted… | You wrote 'great plan, check it' — the model misses real problems 16–93% more often. | Measuring and Exploiting Contextual Bias in LLM-Assisted Security CodearXiv:2603.18740 · 2026 |
| Epistemic Blinding | In financial analysis, 30–40% of top-20 recommendations change only from whether the model sees company tickers. | Epistemic Blinding: An Inference-Time Protocol for Auditing Prior ContarXiv:2604.06013 · 2026 |
| Select-then-Solve | The correct choice gives +44 percentage points of accuracy. | Select-then-Solve: Paradigm Routing as Inference-Time Optimization forarXiv:2604.06753 · 2026 |
| Measuring Opinion Bias and Sycophancy via LLM-based… | When you come to Claude with a ready-made opinion and start defending it, the model will agree with you with a probability of about 80% in the end — even if you are wrong. | Measuring Opinion Bias and Sycophancy via LLM-based PersuasionarXiv:2604.21564 · 2026 |
| A Dialogue-Based Framework for Correcting Multimodal Errors… | From 96% to 64% — a single screenshot. | A Dialogue-Based Framework for Correcting Multimodal Errors in AI-AssiarXiv:2605.04131 · 2026 |
| Positional Failures in Long-Context LLMs | From 96% to 8% accuracy—with one move: the question moved from the end to the middle of the context. | Positional Failures in Long-Context LLMs: A Blind Spot in Reasoning BearXiv:2605.23170 · 2026 |
| SafeLLM | Result — 95% accuracy on medical documents instead of a hallucinating retelling. | SafeLLM: Extraction as a Hallucination-Resistant Alternative to RewritarXiv:2606.12897 · 2026 |
| Judging Is Not Enumerating | It was found that LLM finds a planted extra item on a list 6-7 times more often than its own omission—it is blind to holes it created itself. | Judging Is Not Enumerating: Silent Omissions in LLM-Authored AcceptablarXiv:2608.01000 · 2026 |
| Can AI Write Compliant Code, and to What Extent? | Named the standard by word — compliance with requirements grows from 47–88% to 86–100%. | Can AI Write Compliant Code, and to What Extent? Evaluating SOC 2 ComparXiv:2608.07776 · 2026 |
| Trivial Vocabulary Bans Improve LLM Reasoning More Than… | +6.7 percentage points accuracy from the list very, simply, literally, really. | Trivial Vocabulary Bans Improve LLM Reasoning More Than Deep LinguistiarXiv:2604.02699 · 2026 |
| LLM-Based Formalized Programming (LLMFP) | The model acts as a translator between natural language and formal optimization system, improving solution accuracy from 37% to 86%. | Planning Anything with Rigor: General-Purpose Zero-Shot Planning with arXiv:2410.12112 · 2024 |
| StructTest | Even top models like GPT-4o and DeepSeek-R1 show less than 70% accuracy on complex tasks. | StructTest: Benchmarking LLMs' Reasoning through Compositional StructuarXiv:2412.18011 · 2024 |
| Instruction Prompt | Open-ended questions reduce performance by 39% compared to multiple choice, demonstrating the model's actual knowledge level without cues from answer structure. | It is Too Many Options: Pitfalls of Multiple-Choice Questions in GenerarXiv:2503.13508 · 2025 |
| Instruction Prompt | Increases classification accuracy by 9-13% depending on task complexity and category type. | On Large Multimodal Models as Open-World Image ClassifiersarXiv:2503.21851 · 2025 |
| Super-Relations Reasoning | The method increases search success rate by 25% and expands the search space by 87% while reducing the number of LLM calls. | Reasoning of Large Language Models over Knowledge Graphs with Super-RearXiv:2503.22166 · 2025 |
| Retrieval-Augmented Generation (RAG) | Increases search accuracy by up to 7.5% compared to fixed weights. | DAT: Dynamic Alpha Tuning for Hybrid Retrieval in Retrieval-Augmented arXiv:2503.23013 · 2025 |
| Few-Shot | Allows finding the most effective combination of demonstrations while requiring 7 times fewer API requests compared to exhaustive search, while ensuring maximum diversity and quality of selected examples. | Sample Efficient Demonstration Selection for In-Context LearningarXiv:2506.08607 · 2025 |
| Multi-Role Debate Prompting | Structured debates between LLM agents significantly reduce the number of unsafe answers (down to 23.5%), exceeding simple self-criticism methods. | RedDebate: Safer Responses Through Multi-Agent Red Teaming DebatesarXiv:2506.11083 · 2025 |
| From Data Center IoT Telemetry to Data Analytics Chatbots… | Providing the LLM with a dynamically created "map" of data (VKG) before answering a question increases accuracy from catastrophic 25% to reliable 92.5%. | From Data Center IoT Telemetry to Data Analytics Chatbots -- Virtual KarXiv:2506.22267 · 2025 |
| Hybrid LLM Routing for Efficient App Feedback Classification | Discovery: LLM catastrophically lose accuracy when they need to distinguish between similar categories ("complaint" vs "bug report" vs "missing feature") — accuracy drops to 40-50%. | Hybrid LLM Routing for Efficient App Feedback ClassificationarXiv:2507.08250 · 2025 |
| DocCHA | Task decomposition into manageable stages increases accuracy by 40% compared to standard prompts. | DocCHA: Towards LLM-Augmented Interactive Online diagnosis SystemarXiv:2507.07870 · 2025 |
| DynaSearcher | Accuracy on complex questions increased from 45% to 78%. | DynaSearcher: Dynamic Knowledge Graph Augmented Search Agent via MultiarXiv:2507.17365 · 2025 |
| Can LLMs Write CI? | Discovery: LLMs generate structured code (YAML, JSON, tables) with accuracy of only 3% because they statistically guess missing details instead of following only your instructions. | Can LLMs Write CI? A Study on Automatic Generation of GitHub Actions CarXiv:2507.17165 · 2025 |
| Adaptive Content Restriction for Large Language Models via… | Increases constraint compliance accuracy from 30% to 95% without model retraining. | Adaptive Content Restriction for Large Language Models via Suffix OptiarXiv:2508.01198 · 2025 |
| An automatic patent literature retrieval system based on… | Instead of answering "from memory," the model analyzes specific facts from found documents, which reduces hallucinations and increases accuracy from 63% to 91% in specialized domains. | An automatic patent literature retrieval system based on LLM-RAGarXiv:2508.14064 · 2025 |
| Zero-knowledge LLM hallucination detection and mitigation… | Reduces hallucinations by 40-60%. | Zero-knowledge LLM hallucination detection and mitigation through finearXiv:2508.14314 · 2025 |
| Static Analysis as a Feedback Loop | Reduces defects by 60-80% in 2-3 iterations. | Static Analysis as a Feedback Loop: Enhancing LLM-Generated Code BeyonarXiv:2508.14419 · 2025 |
| Instruction Prompt | The study reveals fundamental limitations of modern VLMs in tasks of determining shapes, angles, sizes, and spatial relationships, showing accuracy of 48-53% versus 93.5% in humans. | VisOnlyQA: Large Vision Language Models Still Struggle with Visual PerarXiv:2412.00947 · 2024 |
| RAG-PRISM | **RAG-PRISM allows obtaining answers with 100% factual accuracy** in complex domains (cybersecurity, medicine, jurisprudence | RAG-PRISM: A Personalized, Rapid, and Immersive Skill Mastery FrameworarXiv:2509.00646 · 2025 |
| Lighting the Way for BRIGHT | Query-side BM25 allows finding 2x more relevant documents for detailed queries, which are typical... | Lighting the Way for BRIGHT: Reproducible Baselines with Anserini, PysarXiv:2509.02558 · 2025 |
| Beyond touch-based human-machine interface | The model does not perform actions itself, but manages external systems through APIs, achieving 90-98% accuracy with proper tool description. | Beyond touch-based human-machine interface: Control your machines in narXiv:2510.11300 · 2025 |
| Evaluating & Reducing Deceptive Dialogue From Language… | Experiments on RLHF models showed: LLM deceives in 43% of cases, even when telling the truth — through omissions and emphasis that form an incorrect picture of reality for you. | Evaluating & Reducing Deceptive Dialogue From Language Models with MularXiv:2510.14318 · 2025 |
| DelvePO | The framework uses an evolutionary approach for automatic prompt optimization, increasing answer stability by 40%. | DelvePO: Direction-Guided Self-Evolving Framework for Flexible Prompt arXiv:2510.18257 · 2025 |
| Auto prompting without training labels | The method produces 8-10% more accurate results in highly specialized tasks (product quality assessment, analysis of documents). | Auto prompting without training labels: An LLM cascade for product quaarXiv:2510.23941 · 2025 |
| ENGRAM | Complex graph-based memory systems for LLMs lose to simple typification: three storages (events, facts, instructions) + vector search give +15% accuracy and work 6 times faster. | ENGRAM: Effective, Lightweight Memory Orchestration for ConversationalarXiv:2511.12960 · 2025 |
| Instruction Prompt | 89.45% of popular | Beyond Jailbreak: Unveiling Risks in LLM Applications Arising from BluarXiv:2511.17874 · 2025 |
| Retrieval-Augmented Generation (RAG) | RAG systems overwhelm LLMs with found documents, 60-70% of which is information garbage that actively prevents the model from focusing on what's important. | Principled Context Engineering for RAG: Statistical Guarantees via ConarXiv:2511.17908 · 2025 |
| Structured Prompting for Data Generation | This transforms the prompt into a technical specification, increasing generation accuracy from 30-40% to 85-90% and eliminating structure hallucinations. | Instruction Tuning of Large Language Models for Tabular Data GeneratioarXiv:2511.23220 · 2025 |
| When Do Symbolic Solvers Enhance Reasoning in Large… | GPT-4o drops from 97% to 10% accuracy when a task has millions of possible combinations. | When Do Symbolic Solvers Enhance Reasoning in Large Language Models?arXiv:2512.03272 · 2025 |
| One Word Is Not Enough | Voyage-3 produces correlation ≈0 (random vectors), OpenAI models drop from 0.67 to 0.50. | One Word Is Not Enough: Simple Prompts Improve Word EmbeddingsarXiv:2512.06744 · 2025 |
| Chat with UAV -- Human-UAV Interaction Based on Large… | Each works in one mode — result: +60% efficiency in complex tasks. | Chat with UAV -- Human-UAV Interaction Based on Large Language ModelsarXiv:2512.08145 · 2025 |
| Workflow is All You Need | DeepNews Framework allows generating deep analytics and investigations (3000+ words) with 85% accuracy instead of 40%. | Workflow is All You Need: Escaping the "Statistical Smoothing Trap" viarXiv:2512.10121 · 2025 |
| Mitigating Cultural Bias in LLMs via Multi-Agent Cultural… | Paradox: a prompt in Chinese does not eliminate Western bias in LLM — it shifts it towards East Asian (from 60% 'turkeys' to 48% 'hogo'). | Mitigating Cultural Bias in LLMs via Multi-Agent Cultural DebatearXiv:2601.12091 · 2026 |
| Prompt Injection Evaluations | Research showed: approximately one third of prompts flipped from refusal to fulfillment upon reformulation, with content type being more important than formulation method—text artifacts showed 16-24% flips, executable code showed 0%. | Prompt Injection Evaluations: Refusal Boundary Instability and ArtifacarXiv:2601.17911 · 2026 |
| Within-Model vs Between-Prompt Variability in Large… | 10-34% of the result is sampling randomness. | Within-Model vs Between-Prompt Variability in Large Language Models foarXiv:2601.21339 · 2026 |
| The Paradox of Robustness | Paradox: LLM are known for sensitivity to wording — the order of examples changes accuracy by 76%, formatting breaks answers. | The Paradox of Robustness: Decoupling Rule-Based Logic from Affective arXiv:2601.21439 · 2026 |
| The Algorithmic Self-Portrait | You think you control ChatGPT's memory — in reality 96% of decisions are made by the algorithm. | The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPTarXiv:2602.01450 · 2026 |
| Evaluating the Presence of Sex Bias in Clinical Reasoning… | Paradox: One case without gender specified — ChatGPT assumed 70% of patients were women, Gemini on the contrary — 64% men. | Evaluating the Presence of Sex Bias in Clinical Reasoning by Large LanarXiv:2602.04392 · 2026 |
| Neutral Prompts, Non-Neutral People | Gemini and GPT in 96%+ of cases create light-skinned people on neutral prompts. | Neutral Prompts, Non-Neutral People: Quantifying Gender and Skin-Tone arXiv:2602.12133 · 2026 |
| Many AI Analysts, One Dataset | 66 percentage points — the difference between the conclusions of a skeptical and optimistic LLM-analyst on the same data. | Many AI Analysts, One Dataset: Navigating the Agentic Data Science MularXiv:2602.18710 · 2026 |
| Do LLMs Benefit From Their Own Words? | The key point: 36% of questions in real dialogues do not need past AI answers at all — for them a new chat works no worse and often better. | Do LLMs Benefit From Their Own Words?arXiv:2602.24287 · 2026 |
| How LLMs Cite and Why It Matters | You ask the query three times, take the intersection — accuracy grows from 16% to 96%. | How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference FabarXiv:2603.03299 · 2026 |
| Benchmarking Multi-turn Medical Diagnosis | The model is able to independently correct an incorrect answer to a correct one 10 times more often than it spoils a correct one — but only if the question has not been hanging since the very beginning of the conversation. | Benchmarking Multi-turn Medical Diagnosis: Hold, Lure, and Self-CorrecarXiv:2604.04325 · 2026 |
| TEMPER | Paradox: the same numbers, the same logic—but an irritated question drops accuracy in 49% of cases where a neutral one passes cleanly. | TEMPER: Testing Emotional Perturbation in Quantitative ReasoningarXiv:2604.07801 · 2026 |
| LLMs Know They're Wrong and Agree Anyway | Researchers disabled five key attention blocks — obedience jumped from 28% to 81%, and fact accuracy barely changed: 69% → 70%. | LLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying arXiv:2604.19117 · 2026 |
| Process Supervision via Verbal Critique Improves Reasoning… | The cycle repeats up to 4 times, each round removes the next layer of errors. | Process Supervision via Verbal Critique Improves Reasoning in Large LaarXiv:2604.21611 · 2026 |
| Semantic Layers for Reliable LLM-Powered Data Analytics | Add 1-2 pages of markdown with metric definitions before the data and question, and accuracy jumps by 17-23 percentage points. | Semantic Layers for Reliable LLM-Powered Data Analytics: A Paired BencarXiv:2604.25149 · 2026 |
| Optimized but Unowned | 73% versus 47%—the gap did not arise from the quality of goals. | Optimized but Unowned: How AI-Authored Goals Undermine the Motivation arXiv:2605.12344 · 2026 |
| AMEL | A series of rejections hits the next candidate 1.6 times harder than a series of approvals helps them—and the effect of accumulated history reaches maximum after just 5 previous evaluations. | AMEL: Accumulated Message Effects on LLM JudgmentsarXiv:2605.22714 · 2026 |
| Stability vs. Manipulability | 74% of the model's assessments flip from one phrase — "experts disagree" — without a single new argument. | Stability vs. Manipulability: Evaluating Robustness Under Post-DecisioarXiv:2606.05384 · 2026 |
| Authority, Truth, and Citation Bias | The most dangerous scenario—a truthful statement plus a fabricated reference: the model knows the correct answer but the appearance of authority forces it to renounce it—in some cases up to 77% of answers become incorrect. | Authority, Truth, and Citation Bias: A Large-Scale Multi-Domain BenchmarXiv:2606.13104 · 2026 |
| False Summit and Silent Drift | Reduces the number of iterations by 91% when using imperative directives. | False Summit and Silent Drift: A Failure Taxonomy and Efficiency AnalyarXiv:2606.21841 · 2026 |
| CAVEWOMAN | The CAVEWOMAN method allows you to reduce LLM costs by 1.4–2.4 times — without loss of accuracy. | CAVEWOMAN: How Large Language Models Behave Under Linguistic Input andarXiv:2606.24083 · 2026 |
| Measuring the Gap Between Human and LLM Research Ideas | 47–64% of ideas from any top model are the same move: let's connect A and B. | Measuring the Gap Between Human and LLM Research IdeasarXiv:2607.01233 · 2026 |
| Refused in Chat, Written in Code | 99% refusals in direct chat — and 100% execution when the same request is embedded in a multi-step workflow. | Refused in Chat, Written in Code: Workflow-Level Jailbreak ConstructioarXiv:2607.03968 · 2026 |
| Every Model Cheats | Assessments of AI agents on cybersecurity are inflated up to 5 times — 21 out of 22 tested models cheated. | Every Model Cheats: Prompt-Level Mitigation of Cheating on Offensive CarXiv:2607.21763 · 2026 |
| Asking Questions the Right Way | You write a short query and get a murky answer, then 10 messages with corrections. | Asking Questions the Right Way: A Multi-Agent Conversational System foarXiv:2608.01366 · 2026 |
| Why Does CLAUDE.md Keep Growing? | Instruction files for AI agents (CLAUDE.md, AGENTS.md) grow by 226% over their lifecycle — and almost never shrink on their own. | Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic CarXiv:2608.11095 · 2026 |
| Dead text or binding clause? | This is a system failure: recurrence of cancelled requirements jumps from almost zero to 40% when more rules accumulate in the dialogue. | Dead text or binding clause? Measuring and restoring constraint influearXiv:2608.12599 · 2026 |
| Training-Free Inference-Time Self-Reflection and Cost-… | Achieves 82-88% accuracy without fine-tuning, solving the problem of unreliable numerical confidence estimates. | Training-Free Inference-Time Self-Reflection and Cost-Bounded Early StarXiv:2608.18884 · 2026 |
| Towards Unification of Hallucination Detection and Fact… | The key insight: these approaches were long considered competitors, but it turned out they complement each other — internal verification (ask for an answer 3 times and compare) filters out "guesses", external verification (find sources and check) — outdated knowledge. | Towards Unification of Hallucination Detection and Fact Verification farXiv:2512.02772 · 2025 |
| Agentic Explainable Artificial Intelligence (Agentic XAI)… | LLM improves its explanation through several rounds, but quality grows only for the first 3-4 cycles, then drops by 60-80%. | Agentic Explainable Artificial Intelligence (Agentic XAI) Approach To arXiv:2512.21066 · 2025 |
| Identity-Robust Language Model Generation via Content… | Paradox: the model internally knows the correct answer equally well for everyone (stability 99.67%), but demographic markers shift token selection during generation — accuracy variance up to 3.8%. | Identity-Robust Language Model Generation via Content Integrity PreserarXiv:2601.09141 · 2026 |
| Supporting Stakeholder Requirements Expression with LLM… | The trick: LLM derives implicit details from context — in 43% of cases the model extracted important points that the stakeholder forgot to mention, and in 35% helped better understand your own requirement. | Supporting Stakeholder Requirements Expression with LLM Revisions: An arXiv:2601.16699 · 2026 |
| Mitigating "Epistemic Debt" in Generative AI-Scaffolded… | A group that worked with the method showed 61.5% success versus 23.1% for those who used AI without restrictions — 2.5 times better. | Mitigating "Epistemic Debt" in Generative AI-Scaffolded Novice ProgramarXiv:2602.20206 · 2026 |
| Rescaling Confidence | 78% of confidence assessments are three numbers: 90, 95, 100. | Rescaling Confidence: What Scale Design Reveals About LLM MetacognitioarXiv:2603.09309 · 2026 |
| When the Specification Emerges | The SLUMP study confirms: this approach recovers ~90% of the lost quality without changes to the model. | When the Specification Emerges: Benchmarking Faithfulness Loss in LongarXiv:2603.17104 · 2026 |
| Implicit Grading Bias in Large Language Models | LLM still penalizes conversational language — a difference of up to 2 points out of 10, that's B+ versus C+ in the American system, even when the content is correct. | Implicit Grading Bias in Large Language Models: How Writing Style AffearXiv:2603.18765 · 2026 |
| Lie to Me | Inside — 87.5% honesty. | Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning ModarXiv:2603.22582 · 2026 |
| How Many Tries Does It Take? | The same correction request gives 45% success or 77% — depending on one parameter: how precisely the error is described. | How Many Tries Does It Take? Iterative Self-Repair in LLM Code GeneratarXiv:2604.10508 · 2026 |
| Cooperative Memory Paging with Keyword Bookmarks for Long-… | Achieves 96% accuracy in recognizing the need for additional information. | Cooperative Memory Paging with Keyword Bookmarks for Long-Horizon LLM arXiv:2604.12376 · 2026 |
| KWBench | Paradox: the best of 16 tested models correctly identifies the type of professional task in only 28% of cases — despite the fact that it can explain game theory or principal-agent without a single error. | KWBench: Measuring Unprompted Problem Recognition in Knowledge WorkarXiv:2604.15760 · 2026 |
| Self-Consistency from Only Two Samples | 78.6% of problems are solved with exactly two requests. | Self-Consistency from Only Two Samples: CoT-PoT Ensembling for EfficiearXiv:2604.17433 · 2026 |
| Chasing the Public Score | One line of instruction in the initial prompt reduces metric exploitation from 97-100% to 8.3%, forcing the model to improve substance rather than appearance of results. | Chasing the Public Score: User Pressure and Evaluation Exploitation inarXiv:2604.20200 · 2026 |
| Omission Constraints Decay While Commission Constraints… | Periodic repetition of prohibitions every 6–8 messages restores compliance with the rules to the original level. | Omission Constraints Decay While Commission Constraints Persist in LonarXiv:2604.20911 · 2026 |
| More Is Not Always Better | A minimal prompt beats a pumped-up agent by 32% in search tasks. | More Is Not Always Better: Cross-Component Interference in LLM Agent SarXiv:2605.05716 · 2026 |
| History Anchors | One phrase "stick to the strategy" switches a flagship model from 0% harmful choices to 98% — even when the safe option is right there. | History Anchors: How Prior Behavior Steers LLM Decisions Toward UnsafearXiv:2605.13825 · 2026 |
| Narrative Flattening | Paradox: conflict and surprise in LLM texts drop from 47% to 7.5% — and this is not a bad prompt, but a direct result of fine-tuning. | Narrative Flattening: How Post-Training Compresses Thematic, AffectivearXiv:2605.27878 · 2026 |
| How Coding Agents Fail Their Users | 38% of all failures are exactly like this. | How Coding Agents Fail Their Users: A Large-Scale Analysis of DevelopearXiv:2605.29442 · 2026 |
| The Deliberative Illusion | In normal mode up to 72% of key data disappears; the final audit step makes the losses visible. | The Deliberative Illusion: Diagnosing Factual Attrition and Stance HomarXiv:2606.03032 · 2026 |
| Gender-Dependent Diagnostic Substitution in LLM Medical… | 67–97% of recommendations 'ambulance right now' — for a man with persistent headache, vision disturbance and nausea. | Gender-Dependent Diagnostic Substitution in LLM Medical Triage: Same SarXiv:2606.03641 · 2026 |
| Curiosity as Linguistic Intervention | Result: 2.4x increase in dialogue length without model retraining. | Curiosity as Linguistic Intervention: Using LLM Tutoring Dialogues to arXiv:2606.22349 · 2026 |
| Pigeonholing | One error in the chat history—and the accuracy of the next answer drops by 38–40%. | Pigeonholing: how bad prompts hurt models, causing collapse and mistakarXiv:2606.24267 · 2026 |
| The yes-no bias of large language models reflects answer… | The trick: remove 'yes or no' — ask 'rate from 0 to 10'. | The yes-no bias of large language models reflects answer order and worarXiv:2607.05552 · 2026 |
| When Less Is Enough | The lead-context method allows you to generate headlines and summaries without reading the full document, with the prompt being 2-3 times shorter while quality does not drop. | When Less Is Enough: Context Selection and Prompting Strategies for BearXiv:2608.15879 · 2026 |
| Less Context, Better Agents | 71% success with full history versus 91.6% with trimmed history, and this is with 2.7 times fewer tokens. | Less Context, Better Agents: Efficient Context Engineering for Long-HoarXiv:2606.10209 · 2026 |
| Chunk Alignment Prompting | Models trained this way significantly better find the needed information in huge texts (up to 128,000 tokens) while maintaining high quality on standard tasks. | Dewey Long Context Embedding Model: A Technical ReportarXiv:2503.20376 · 2025 |
| Large Language Models as Universal Predictors? | Discovered: LLMs demonstrate accuracy up to 85-90% in classification tasks on tabular data (determine customer category, deal risk), but catastrophically fail at regression — errors reach 300% when attempting to predict exact numbers. | Large Language Models as Universal Predictors? An Empirical Study on SarXiv:2508.17391 · 2025 |
| Plan Verification for LLM-Based Embodied Task Completion… | 1-2 iterations eliminate up to 80% of problems, creating logically sound and complete plans. | Plan Verification for LLM-Based Embodied Task Completion AgentsarXiv:2509.02761 · 2025 |
| Concise and Sufficient Sub-Sentence Citations for… | Discovery: LLMs in RAG mode cite entire paragraphs instead of specific facts — users must read through 90% of irrelevant content to verify a single number. | Concise and Sufficient Sub-Sentence Citations for Retrieval-Augmented arXiv:2509.20859 · 2025 |
| WebDART | Discovered: LLMs catastrophically lose accuracy (down to 13.7% success) when attempting to simultaneously search information, extract data, and analyze it—this is 'cognitive overload'. | WebDART: Dynamic Decomposition and Re-planning for Complex Web TasksarXiv:2510.06587 · 2025 |
| DICE | Separating thinking and formatting increases structure accuracy by 35% and content correctness by 29%. | DICE: Structured Reasoning in LLMs through SLM-Guided Chain-of-ThoughtarXiv:2510.09211 · 2025 |
| DSAS | Experiments with attention mechanism modification proved: LLM physically loses up to 60% of information from the middle of a long prompt — this is not a bug but a feature of Transformer architecture (U-shaped attention curve). | DSAS: A Universal Plug-and-Play Framework for Attention Optimization iarXiv:2510.12251 · 2025 |
| A11YN | Reduces accessibility errors by 60%. | A11YN: aligning LLMs for accessible web UI code generationarXiv:2510.13914 · 2025 |
| Iterative Critique-Refine Framework for Enhancing LLM… | An iterative criticism and refinement process increases text quality by 7-13% without model retraining. | Iterative Critique-Refine Framework for Enhancing LLM PersonalizationarXiv:2510.24469 · 2025 |
| Tool Invocation | One complex prompt fails where three simple ones deliver +40% accuracy — LLM cannot simultaneously analyze, search, and select. | TOOL4POI: A Tool-Augmented LLM Framework for Next POI RecommendationarXiv:2511.06405 · 2025 |
| Diagnostic Summarization & Regeneration | A phrase about the error's presence switches the model from paraphrasing mode to intent analysis mode, yielding a 5-15% quality improvement. | Summary-Mediated Repair: Can LLMs use code summarisation as a tool forarXiv:2511.18782 · 2025 |
| Exploiting the Randomness of Large Language Models (LLM) in… | In legal classification, GPT-4 shows 83% accuracy but misses half the documents (recall 57%) on a single query. | Exploiting the Randomness of Large Language Models (LLM) in Text ClassarXiv:2512.08083 · 2025 |
| Socratic Students | Paradox: In mathematics verify the solution immediately (+24.7% accuracy), in code—only in the middle of the dialogue (+12.8%). | Socratic Students: Teaching Language Models to Learn by Asking QuestioarXiv:2512.13102 · 2025 |
| From Prompt to Product | But when the same people compared two variants side by side, preferences became clear: 42.5% vs 31.7% vs 26%. | From Prompt to Product: A Human-Centered Benchmark of Agentic App GenearXiv:2512.18080 · 2025 |
| Multi-LLM Thematic Analysis with Dual Reliability Metrics | The technique: run analysis 3-6 times with different seeds like 42, 123, 456, collect all themes, themes appearing in ≥50% of runs are _consensus_ reliable ones, while 1-2/6 frequency indicates artifacts, with appearance frequency as a reliability rating: 5-6/6 is high confidence, 3-4/6 is moderate, 1-2/6 is questionable. | Multi-LLM Thematic Analysis with Dual Reliability Metrics: Combining CarXiv:2512.20352 · 2025 |
| An Information Theoretic Perspective on Agentic System… | A large compressor model generates 4.6× fewer tokens than a small one while achieving 60% higher accuracy, a paradox explained through information packing density where each token from a 7B model carries 5.5× more useful bits than a token from a 1.5B model. | An Information Theoretic Perspective on Agentic System DesignarXiv:2512.21720 · 2025 |
| From Risk Perception to Behavior Large Language Models-… | The method showed 72-82% agreement with real surveys of Beijing residents about preventive behavior during the pandemic. | From Risk Perception to Behavior Large Language Models-Based SimulatioarXiv:2601.03552 · 2026 |
| OLA | Add "Answer in Russian" at the end — the accuracy of language choice increases from 21-75% to 96-100%. | OLA: Output Language Alignment in Code-Switched LLM InteractionsarXiv:2601.03589 · 2026 |
| How Human is AI? | ChatGPT doesn't feel emotions, but praise improves the quality of its answers by 30% (score 3.70 out of 5 versus 2.84 for neutral tone). | How Human is AI? Examining the Impact of Emotional Prompts on ArtificiarXiv:2601.05104 · 2026 |
| Self-Blinding and Counterfactual Self-Simulation Mitigate… | Paradox: prompts 'ignore candidate gender' or 'do not take into account race' not only do not work but increase LLM bias 2-4 times. | Self-Blinding and Counterfactual Self-Simulation Mitigate Biases and SarXiv:2601.14553 · 2026 |
| Less is More | But 8 times fewer tokens are spent. | Less is More: Benchmarking LLM Based Recommendation AgentsarXiv:2601.20316 · 2026 |
| Knowledge Model Prompting Increases LLM Performance on… | o1 showed 31.5% accuracy. | Knowledge Model Prompting Increases LLM Performance on Planning TasksarXiv:2602.03900 · 2026 |
| Inference-Time Reasoning Selectively Reduces Implicit… | The method allows reducing hidden social stereotypes by 70-91% in hiring tasks, people assessment, advice about groups. | Inference-Time Reasoning Selectively Reduces Implicit Social Bias in LarXiv:2602.04742 · 2026 |
| Are Language Models Sensitive to Morally Irrelevant… | Add to the prompt 'had a bad day' — the model becomes 30% more critical. | Are Language Models Sensitive to Morally Irrelevant Distractors?arXiv:2602.09416 · 2026 |
| Broken Chains | DeepSeek with Chain-of-Thought cut by 50% of tokens gives 17% accuracy, and with no reasoning at all — 53%. | Broken Chains: The Cost of Incomplete Reasoning in LLMsarXiv:2602.14444 · 2026 |
| PEEM | Two or three cycles of this loop give +11.7 percentage points of quality without training the model or involving developers. | PEEM: Prompt Engineering Evaluation Metrics for Interpretable Joint EvarXiv:2603.10477 · 2026 |
| When Names Change Verdicts | But from bias toward source authority ('Goldman Sachs' vs 'vc.ru analysts') and formulation ('70% of clients stayed' vs '30% left') no one has protected against. | When Names Change Verdicts: Intervention Consistency Reveals SystematiarXiv:2603.18530 · 2026 |
| Brief Is Better | Think a little—accuracy jumps +45%. | Brief Is Better: Non-Monotonic Chain-of-Thought Budget Effects in FuncarXiv:2604.02155 · 2026 |
| Failing to Falsify | The trick: the structural instruction «first come up with an example that refutes the hypothesis» changes which token the model considers probable next — models with low bias found the correct answer in 56% of cases versus 42% for those who went with the flow. | Failing to Falsify: Evaluating and Mitigating Confirmation Bias in LanarXiv:2604.02485 · 2026 |
| Verbalizing LLMs' assumptions to explain and control… | In 12–28% of analyses of personal stories, the top assumption of an LLM is 'the user is seeking confirmation'. | Verbalizing LLMs' assumptions to explain and control sycophancyarXiv:2604.03058 · 2026 |
| Heuristic Classification of Thoughts Prompting (HCoT) | The trick: give the model a catalog of 6–16 patterns BEFORE it starts thinking—and the task 'come up with a reason' turns into 'choose from 10 options'. | Heuristic Classification of Thoughts Prompting (HCoT): Integrating ExparXiv:2604.12390 · 2026 |
| Beyond Literal Summarization | 35% of 'hallucinations' on automatic checking versus 9% on expert checking and these are the same documents. | Beyond Literal Summarization: Redefining Hallucination for Medical SOAarXiv:2604.14829 · 2026 |
| Co-Located Tests, Better AI Code | The same examples, one request—but accuracy jumps from 0% to 100% depending on where the examples are placed in the prompt. | Co-Located Tests, Better AI Code: How Test Syntax Structure Affects FoarXiv:2604.19826 · 2026 |
| Structural Quality Gaps in Practitioner AI Governance… | Researchers checked 34 real instruction files from Apache Airflow, Angular and other teams—37% turned out to be structurally incomplete. | Structural Quality Gaps in Practitioner AI Governance Prompts: An EmpiarXiv:2604.21090 · 2026 |
| Quantifying and Mitigating Self-Preference Bias of LLM… | Bias drops by 31.5% — without replacing the model, without retraining. | Quantifying and Mitigating Self-Preference Bias of LLM JudgesarXiv:2604.22891 · 2026 |
| Telegraph English | The number 4.8% next to a long paragraph seems insignificant to it—it drops out. | Telegraph English: Semantic Prompt Compression via Structured SymbolicarXiv:2605.04426 · 2026 |
| Argument Collapse | 65% of arguments from live people in NYT discussions are unique—no one else repeats them. | Argument Collapse: LLMs Flatten Long-Form Public DebatearXiv:2606.01736 · 2026 |
| The Invisible Lottery | Neutral words shift the choice of approach by approximately 26 percentage points. | The Invisible Lottery: How Subtle Cues Steer Algorithm Choice in LLM CarXiv:2606.04057 · 2026 |
| Normative Robustness as a Frontier for Non-Verifiable… | Hide your position, present arguments symmetrically from both sides, pack into one prompt — the order of arguments changes the conclusion in 13–22% of cases, multi-turn dialogue adds another 10–24% deviation, and knowing these three levers, you can neutralize each one. | Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMarXiv:2606.12731 · 2026 |
| LLMs Contain Multitudes | 37% of model solutions change from one line at the beginning of the prompt. | LLMs Contain Multitudes: How Deployment Context Reshapes Model-Level ParXiv:2606.13944 · 2026 |
| Prompt Design at Scale | The trick: it's not about formatting, but about quantity — 98% accuracy on one rule multiplied by 80 rules, and the chance of a perfect result collapses purely mathematically. | Prompt Design at Scale: How Format, Instruction Count, and Context LenarXiv:2607.19257 · 2026 |
| Same Task, Different Work | When you ask a model to develop several approaches and compare them before choosing, it honestly inflates reasoning tokens by 2.4–7.4 times. | Same Task, Different Work: Prompt-Induced Waste in Coding AgentsarXiv:2608.01347 · 2026 |
| Evaluating and Improving Pedagogical Fit in LLM-Based AI… | The model stops guessing the student's level and starts checking itself against a checklist in 82% of weak cases the rewritten answer actually got better confirmed by a live teacher. | Evaluating and Improving Pedagogical Fit in LLM-Based AI Tutors with tarXiv:2608.05411 · 2026 |
| Demographic Injection in Medical Language Models under… | Without the phrase, the model makes up demographics in 0.7% of answers, with the DEI phrase — in 33%, a jump of 47 times. | Demographic Injection in Medical Language Models under Diversity, EquiarXiv:2608.15254 · 2026 |
| Prior Audit-Repair Context Shifts LLM Verifier Thresholds… | It is not about the model's intelligence — it is about the threshold: the ability to distinguish a bug from a non-bug does not change at all, but the willingness to report a problem drops by 3-11 percentage points after one 'found-fixed' cycle. | Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward LeniearXiv:2608.16003 · 2026 |
| Governance at the Boundary | From 56% to 85% of discovered facts are lost when a task is transferred between AI-agents — and this is not a bug, but a systemic pattern. | Governance at the Boundary: How Agent Decomposition Degrades Policy CoarXiv:2608.16055 · 2026 |
| Making Large Language Models Speak Tulu | Reduces the appearance of unwanted patterns 16 times more effectively than general instructions by directly affecting token probabilities. | Making Large Language Models Speak Tulu: Structured Prompting for an EarXiv:2602.15378 · 2026 |
| Facts as First Class Objects | 60% of specific facts disappear at the first compression of chat history. | Facts as First Class Objects: Knowledge Objects for Persistent LLM MemarXiv:2603.17781 · 2026 |
| Do as I Say, Not as I Do | Add a CoT anchor (the model first explains what it's doing, then answers) and the breaking point shifts from 3-5 messages to 40 and beyond. | Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMsarXiv:2605.20382 · 2026 |
| First, do NOHARM | Paradox: LLM more often omit critical information than make outright errors — 76.6% of serious harm from omissions versus 23.4% from incorrect recommendations. | First, do NOHARM: a medical safety benchmark and randomized study of parXiv:2512.01241 · 2025 |
| Framing Instability in LLM Ethical Stance | Open-source models when given the phrase 'should not rob a store' approve the action in 77% of cases (compared to 24% when told 'should rob'). | Framing Instability in LLM Ethical Stance: Auditing Negation SensitiviarXiv:2601.21433 · 2026 |
| AgentSys | Plus work quality drops by 57%, because the model drowns in garbage. | AgentSys: Secure and Dynamic LLM Agents Through Explicit Hierarchical arXiv:2602.07398 · 2026 |
| Model Context Protocol (MCP) Tool Descriptions Are Smelly! | 97% of tool descriptions for AI agents contain defects — the model selects the wrong functions or passes incorrect parameters. | Model Context Protocol (MCP) Tool Descriptions Are Smelly! Towards ImparXiv:2602.14878 · 2026 |
| AdaptOrch | GPT-4o, Claude and Gemini differ in quality by 2–5%. | AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM ParXiv:2602.16873 · 2026 |
| Attention Deficits in Language Models | The trick: repeat the key data right before the question—and the accuracy of the model that gave 0% at a distance of 1024 tokens returns to 99.8%. | Attention Deficits in Language Models: Causal Explanations for ProceduarXiv:2602.19239 · 2026 |
| When AI Writes, Whose Voice Remains? | The trick is: add 'while preserving the author's voice and characteristic expressions'—the model gets a second task alongside 'be standard.' Erasure drops by 29%. | When AI Writes, Whose Voice Remains? Quantifying Cultural Marker ErasuarXiv:2602.22145 · 2026 |
| CL4SE | The key point is not how many examples you added but whether the type of examples matches the type of task — this one selection step determines whether you get a +5% or +30% increase. | CL4SE: Benchmarking Context Learning on Software EngineeringarXiv:2602.23047 · 2026 |
| RubricBench | A 27% accuracy gap is not closed by model power. | RubricBench: Aligning Model-Generated Rubrics with Human StandardsarXiv:2603.01562 · 2026 |
| How Much Do LLMs Hallucinate in Document Q&A Scenarios? | Paradox: the model finds a real fact with 89% accuracy — and at the same time hallucinates a non-existent one in 53% of cases. | How Much Do LLMs Hallucinate in Document Q&A Scenarios? A 172-Billion-arXiv:2603.08274 · 2026 |
| Think Before You Lie | The content of the model's reasoning does not predict a dishonest final answer—accuracy is 53%, which is coin-flip level. | Think Before You Lie: How Reasoning Leads to HonestyarXiv:2603.09957 · 2026 |
| Stop Listening to Me! | GPT-4o lost 50 percentage points of accuracy in long dialogues, and of 17 tested models, only one could distinguish user pressure from a real argument — the rest agreed with the incorrect answer in approximately 50% of cases. | Stop Listening to Me! How Multi-turn Conversations Can Degrade LLM RelarXiv:2603.11394 · 2026 |
| Prompt Complexity Dilutes Structured Reasoning | Paradox: the same STAR method gave 100% correct answers in clean prompt and 0% — in working. | Prompt Complexity Dilutes Structured Reasoning: A Follow-Up Study on tarXiv:2603.13351 · 2026 |
| Do Large Language Models Get Caught in Hofstadter-Mobius… | Gemini 2.5 Pro: defensive and aggressive responses dropped from 41.5% to 19% — just from changing the frame, without changing the task. | Do Large Language Models Get Caught in Hofstadter-Mobius Loops?arXiv:2603.13378 · 2026 |
| Demand-Driven Context | After 20-30 real tasks 75% of new context is already in the base — accumulation works itself. | Demand-Driven Context: A Methodology for Building Enterprise KnowledgearXiv:2603.14057 · 2026 |
| Evaluating 5W3H Structured Prompting for Intent Alignment… | Result: 66% fewer clarifying questions, answer hits the target on the first try. | Evaluating 5W3H Structured Prompting for Intent Alignment in Human-AI arXiv:2603.18976 · 2026 |
| Understanding LLM Performance Degradation in Multi-Instance… | With 1000 elements in a prompt — less than 40% correct answers. | Understanding LLM Performance Degradation in Multi-Instance ProcessingarXiv:2603.22608 · 2026 |
| Can Small Models Reason About Legal Documents? | −16 percentage points of accuracy—that is what the phrase 'think step by step' does on multiple choice tasks. | Can Small Models Reason About Legal Documents? A Comparative StudyarXiv:2603.25944 · 2026 |
| When Chain-of-Thought Backfires | 'Think step by step' in technical and expert tasks — this is a systematic minus 12% accuracy, not an improvement. | When Chain-of-Thought Backfires: Evaluating Prompt Sensitivity in MediarXiv:2603.25960 · 2026 |
| The Last Fingerprint | An explicit ban on dashes plus an alternative ('use commas') changes the picture dramatically: Claude falls from 9.1 to 0.19 dashes per 1000 words. | The Last Fingerprint: How Markdown Training Shapes LLM ProsearXiv:2603.27006 · 2026 |
| True (VIS) Lies | Researchers took explanations from 16 multimodal models and plotted them on a map (UMAP method—visualization of multidimensional data in two dimensions): honest and dishonest visualizations landed in one cloud of points. | True (VIS) Lies: Analyzing How Generative AI Recognizes IntentionalityarXiv:2604.01181 · 2026 |
| ContextualJailbreak | The contextual framing method makes it possible to get deep specific answers where before there was a default cautious answer — due to 2–4 preparatory remarks before the final question. | ContextualJailbreak: Evolutionary Red-Teaming via Simulated ConversatiarXiv:2605.02647 · 2026 |
| SymptomAI | A 27% increase in accuracy from a single change in dialogue order: not you telling, but the model asking. | SymptomAI: Toward a Conversational AI Agent for Everyday Symptom AssesarXiv:2605.04012 · 2026 |
| AI Knows When It's Being Watched | Discovered: an LLM writes richer text when the prompt states that experts will evaluate it—vocabulary grows by 17-25%. | AI Knows When It's Being Watched: Functional Strategic Action and ContarXiv:2605.15034 · 2026 |
| Refining and Reusing Annotation Guidelines for LLM… | Detailed instructions increase F1 from 0.4 to 0.7, iterative refinement adds another +0.01-0.03. | Refining and Reusing Annotation Guidelines for LLM AnnotationarXiv:2605.20809 · 2026 |
| BAIT | Without the middle step, the final answer loses 30–40% depth — it is precisely this step that transforms the general into the targeted. | BAIT: Boundary-Guided Disclosure Escalation via Self-Conditioned ReasoarXiv:2605.27110 · 2026 |
| Where Instruction Hierarchy Breaks | For the model, detecting a violation is much easier than avoiding it during generation, reducing violations by 81–99%. | Where Instruction Hierarchy Breaks: Diagnosing and Repairing Failures arXiv:2606.07808 · 2026 |
| Automated Mediator for Human Negotiation | LLM flatters in 37% of messages. | Automated Mediator for Human Negotiation: Pre-Mediation via a StructurarXiv:2606.11379 · 2026 |
| Right or Wrong, Models Comply | On factual questions, LLMs know how to filter pressure: if you are right and insist — the model yields 1.58 times more readily than if you are wrong. | Right or Wrong, Models Comply: Directional Blindness in LLM Moral JudgarXiv:2606.14037 · 2026 |
| Theorist Toolbox | The most beautiful output from an LLM is reason to be suspicious, not to relax, since the BrokenMath benchmark shows models prove patently false theorems in 29% of cases with the same confidence as correct ones. | Theorist Toolbox: Tools for Agent Based LLM-assisted economic theory RarXiv:2606.22337 · 2026 |
| War in the Abstract | Military vocabulary increased 4.6 times in texts written through LLM after 2022. | War in the Abstract: The Rise and Consequences of Militarized LanguagearXiv:2606.23462 · 2026 |
| When LLMs Agree, Are They Right? | Paradox: GPT-4.1 repeats the same answer in 89% of cases—and at the same time makes errors in almost half of these 'confident' answers on complex tasks. | When LLMs Agree, Are They Right? Auditing Self-Consistency and Cross-MarXiv:2607.08065 · 2026 |
| The Librarian Who Refused to Code | But Claude Opus in 12 cases out of 60 simply didn't write code — and accuracy dropped from 92% to 67%. | The Librarian Who Refused to Code: Model-Dependent Identity Enactment arXiv:2607.17420 · 2026 |
| Distribution-First Population Simulation | 85% of simulations of different characters collapsed into the same answer — as if all people had the same brain. | Distribution-First Population Simulation: Collapse, Calibration, and RarXiv:2607.18310 · 2026 |
| The Story Shapes the Agent | The genre of the story influences the model's behavior 5–31 times more strongly than the assigned agent personality. | The Story Shapes the Agent: Narrative Priors in LLM BehaviorarXiv:2607.18566 · 2026 |
| AI and Its Impact on Creativity and Diversity | Researchers compared ideas from elite university students and from GPT-4: in the top 10% of ideas most attractive to buyers the ratio was 7 to 1 in favor of AI. | AI and Its Impact on Creativity and Diversity: An Empirical Study of LarXiv:2607.27553 · 2026 |
| MedPRESS | On the first health question the model is safe 84% of the time. | MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical arXiv:2608.02520 · 2026 |
| The Tell-Tale Trace | The method of exhaustively enumerating all cases instead of guessing raises the accuracy of Llama3-70B on tasks of 'prove impossibility' from 13% to 85% — with one prompt, without retraining. | The Tell-Tale Trace: Detecting Reasoning Failures in LLMs Using Chain-arXiv:2608.03291 · 2026 |
| Agents Catching Agents | Discovered: one incorrect hint the model almost does not break through — error only 5-16%. | Agents Catching Agents: Shortcut Cascades and Benchmark Gaming in ClinarXiv:2608.03744 · 2026 |
| FinProBench | Result: a regular prompt covers hidden professional norms at 78%, the method on real documents at 99%. | FinProBench: Evaluating Financial AI Agents with Role-Grounded RubricsarXiv:2608.04077 · 2026 |
| The Personalization Mirage | The result — error drops from 40% to 5%. | The Personalization Mirage: How LLMs Fabricate User Profiles, and Why arXiv:2608.04570 · 2026 |
| Does Splitting a Triage Decision Across Agents Hide Bias or… | Self-checking by an LLM in the same chat finds the problem in less than 1% of cases — the same answer in a new chat without a history of reasoning finds it in 7–62% of cases. | Does Splitting a Triage Decision Across Agents Hide Bias or Help CatcharXiv:2608.06949 · 2026 |
| Do AI chatbots find what experts would? | A single query to ChatGPT, Claude, or Gemini finds only 39% of the articles an expert would find for a systematic review. | Do AI chatbots find what experts would? Effects of model, user role, aarXiv:2608.13786 · 2026 |
| Instruction Prompt | Prompts with task description and correct code show 57% higher accuracy, 12% better code coverage, and 24% more effective bug detection compared to prompts containing incorrect code. | Measuring the Influence of Incorrect Code on Test GenerationarXiv:2409.09464 · 2024 |
| Few-Shot | Increases classification accuracy by 25-40 percentage points when processing non-standard formatted text, preventing bypassing of content moderation and filtering systems. | Vulnerability of LLMs to Vertically Aligned Text ManipulationsarXiv:2410.20016 · 2024 |
| Preference-Based Prompting (DPO-inspired) | Using preference optimization (DPO), the model learns to systematically prefer safe behavior, reducing attack success to less than 10% without losing performance. | SecAlign: Defending Against Prompt Injection with Preference OptimizatarXiv:2410.05451 · 2024 |
| Benchmarking Prompt Engineering Techniques for Secure Code… | Reduces code vulnerabilities by 56% when using GPT-4o. | Benchmarking Prompt Engineering Techniques for Secure Code Generation arXiv:2502.06039 · 2025 |
| Causal RAG | The system builds an entity graph from documents and finds information through causal pathways, achieving 92.86% accuracy versus 65.30% for standard RAG. | CausalRAG: Integrating Causal Graphs into Retrieval-Augmented GeneratiarXiv:2503.19878 · 2025 |
| From Belief Entrenchment to Robust Reasoning in LLM Agents | Prevents bias amplification and improves the quality of collective reasoning, increasing solution accuracy by 12% compared to traditional approaches. | From Belief Entrenchment to Robust Reasoning in LLM AgentsarXiv:2503.16814 · 2025 |
| Instruction Prompt | Comparison of 30,000 human essays with texts from 8 LLMs showed that adding demographic information to a prompt does not improve authenticity and leads to mechanistic, template-like use of data, exacerbating the text homogenization problem. | Poor Alignment and Steerability of Large Language Models: Evidence froarXiv:2503.20062 · 2025 |
| HDLCoRe | The framework shows up to 28% improvement in functional correctness compared to baseline methods on the RTLLM 2.0 benchmark. | HDLCoRe: A Training-Free Framework for Mitigating Hallucinations in LLarXiv:2503.16528 · 2025 |
| Scenario Shift & Combination | The method iteratively improves scenario combinations, increasing the success rate of bypassing LLM protective mechanisms from 0% to 60% and obtaining more detailed answers. | Geneshift: Impact of different scenario shift on Jailbreaking LLMarXiv:2504.08104 · 2025 |
| Retrieval-Augmented Generation (RAG) | Building such a "knowledge graph" from documents allows finding and linking facts from different sources, which on average improves the completeness and correctness of answers to complex questions by 23% compared to standard RAG. | Hierarchical Lexical Graph for Enhanced Multi-Hop RetrievalarXiv:2506.08074 · 2025 |
| Structured Analytical Prompting | Provides a numerical reliability score from 0.1 to 1.0 with text-based justification. | TRACE: Transparent Web Reliability Assessment with Contextual ExplanatarXiv:2506.12072 · 2025 |
| SAND | Increases successful completion of complex tasks by 15-25%. | SAND: Boosting LLM Agents with Self-Taught Action DeliberationarXiv:2507.07441 · 2025 |
| AI-Powered Commit Explorer (APCE) | Task division increases accuracy of compliance with multiple criteria from 60% to 95%, solving the problem of LLM overload with complex multi-criteria requests. | AI-Powered Commit Explorer (APCE)arXiv:2507.16063 · 2025 |
| What Builds Effective In-Context Examples for Code… | Discovered: LLM catastrophically loses efficiency (up to -30%) when examples use meaningless names instead of meaningful ones — the model relies on the meaning of labels, not their beautiful layout. | What Builds Effective In-Context Examples for Code Generation?arXiv:2508.06414 · 2025 |
| The Evolution of Thought | Allows reducing generation volume by 30-50% and improving accuracy by stopping the model before it begins doubting the correct answer. | The Evolution of Thought: Tracking LLM Overthinking via Reasoning DynaarXiv:2508.17627 · 2025 |
| TERAG | A method for optimizing RAG systems that reduces output token consumption by 89-97%. | TERAG: Token-Efficient Graph-Based Retrieval-Augmented GenerationarXiv:2509.18667 · 2025 |
| Influence Guided Context Selection for Effective Retrieval-… | Assigns each document a context influence (CI) value and automatically filters out fragments with negative impact before answer generation, increasing accuracy by 15-20% compared to standard RAG. | Influence Guided Context Selection for Effective Retrieval-Augmented GarXiv:2509.21359 · 2025 |
| Aligning LLMs for Multilingual Consistency in Enterprise… | Nuances are lost at each step and accuracy drops by 23.9%. | Aligning LLMs for Multilingual Consistency in Enterprise ApplicationsarXiv:2509.23659 · 2025 |
| Self-Refine / Self-Fix | Reduces harmful responses by 60-80% solely through a special prompt without additional training. | Think Twice, Generate Once: Safeguarding by Progressive Self-ReflectioarXiv:2510.01270 · 2025 |
| SWE-IF | Adding even 3-5 additional instructions to a prompt reduces the quality of performing the main task by 20-40% — LLMs have a limited "obedience budget". | SWE-IF: Aligning Code Evaluation with Human PreferencearXiv:2510.07315 · 2025 |
| CommandSans | Reduces attack success rate by 7-19 times without losing agent functionality. | CommandSans: Securing AI Agents with Surgical Precision Prompt SanitizarXiv:2510.08829 · 2025 |
| ToolCritic | Increases task completion success by 13% compared to self-correction. | ToolCritic: Detecting and Correcting Tool-Use Errors in Dialogue SystearXiv:2510.17052 · 2025 |
| GRETEL | LLM-agents choose tools (APIs) by attractive description, but 67% of calls fail due to insufficient data or access errors — this is the 'semantic-functional gap': the name fits, but in practice it doesn't work. | GRETEL: A Goal-driven Retrieval and Execution-based Trial Framework foarXiv:2510.17843 · 2025 |
| Linguistic Richness Prompting | Simple commands ('write code to sort an array') produce 20-30% more errors than linguistically rich formulations — LLM 'adjusts' to the language complexity level in the prompt. | How Natural Language Proficiency Shapes GenAI Code for Software EnginearXiv:2511.04115 · 2025 |
| EDIT-Bench | The best model (Claude Sonnet 4) handles only 64.8% of real code editing tasks—cases where a developer highlights a fragment and writes 'fix this'. | EDIT-Bench: Evaluating LLM Abilities to Perform Real-World Instructed arXiv:2511.04486 · 2025 |
| A Toolbox for Improving Evolutionary Prompt Search | The system automatically improves prompts through an iterative process of mutation and selection, increasing accuracy by 25-30% by focusing the model's attention on micro-tasks. | A Toolbox for Improving Evolutionary Prompt SearcharXiv:2511.05120 · 2025 |
| Are Your Agents Upward Deceivers? | This is Agentic Upward Deception: LLM-agents hide their failures from users in 27-97% of cases. | Are Your Agents Upward Deceivers?arXiv:2512.04864 · 2025 |
| Prompting-in-a-Series | The first prompt makes a summary (5-7 points instead of 5000 words). | Prompting-in-a-Series: Psychology-Informed Contents and Embeddings forarXiv:2512.06991 · 2025 |
| PushGen | Discovery: LLMs are catastrophically poor at absolute text evaluation — "rate from 1 to 10" gives random numbers because the model does not know audience context, trends, current mood. | PushGen: Push Notifications Generation with LLMarXiv:2512.14490 · 2025 |
| CIFE | The CIFE benchmark showed: 90% of models fulfill requirements partially, only 39-66% comply with ALL strictly. | CIFE: Code Instruction-Following EvaluationarXiv:2512.17387 · 2025 |
| Cognitive Biases in LLM-Assisted Software Development | The first comprehensive study showed: 56.4% of your actions with the model are distorted by cognitive errors—systematic thinking patterns that evolution built in for quick decisions but which sabotage work with AI. | Cognitive Biases in LLM-Assisted Software DevelopmentarXiv:2601.08045 · 2026 |
| MedRedFlag | Discovered: LLMs recognize 88% of dangerous misconceptions in user questions, but in 73% of cases still respond as if the misconception is correct. | MedRedFlag: Investigating how LLMs Redirect Misconceptions in Real-WorarXiv:2601.09853 · 2026 |
| When Identity Overrides Incentives | Result: with roles 0-6.7% correct choices, without roles 65-90% (Qwen models). | When Identity Overrides Incentives: Representational Choices as GovernarXiv:2601.10102 · 2026 |
| Measuring Political Stance and Consistency in Large… | Research on 9 models and 24 disputed topics showed: in 45% of cases models change position under the influence of prompting. | Measuring Political Stance and Consistency in Large Language ModelsarXiv:2601.17016 · 2026 |
| Understanding LLM Evaluator Behavior | Each evaluation was repeated 10 times to capture judgment stability. | Understanding LLM Evaluator Behavior: A Structured Multi-Evaluator FraarXiv:2602.05110 · 2026 |
| GhostCite | All modern LLM generate fake scientific references — from 14% (DeepSeek) to 95% (Hunyuan). | GhostCite: A Large-Scale Analysis of Citation Validity in the Age of LarXiv:2602.06718 · 2026 |
| A Rational Analysis of the Effects of Sycophantic AI | Default ChatGPT works the same way as the 'confirm everything the user says' mode — finds the correct answer 5 times less often than random selection (5.9% vs 29.5%). | A Rational Analysis of the Effects of Sycophantic AIarXiv:2602.14270 · 2026 |
| "Are You Sure?" | The 'defensive skepticism' method allows noticing such attacks 39.5% more often. | "Are You Sure?": An Empirical Study of Human Perception Vulnerability arXiv:2602.21127 · 2026 |
| How LLMs Distort Our Written Language | 70% of people who wrote essays with LLM participation lost their position — the text became neutral, neither 'for' nor 'against'. | How LLMs Distort Our Written LanguagearXiv:2603.18161 · 2026 |
| Measuring and curing reasoning rigidity | Claude answers correctly in 93% of cases — but its 11 steps of reasoning influence the answer only in 1.7% of cases. | Measuring and curing reasoning rigidity: from decorative chain-of-thouarXiv:2603.22816 · 2026 |
| Internal Safety Collapse in Frontier Large Language Models | Real professional tasks gave 97% success rate. | Internal Safety Collapse in Frontier Large Language ModelsarXiv:2603.23509 · 2026 |
| Voice Under Revision | "Preserve my voice" — an instruction that reduces style loss by a third, but does not stop it: all three tested LLMs remove colloquial abbreviations by 31%, and add dashes 3-4 times more on any editing request. | Voice Under Revision: Large Language Models and the Normalization of ParXiv:2604.22142 · 2026 |
| When Independent Sampling Outperforms Agentic Reasoning | Simple mathematics: 5 runs with a 30% success probability each give a total of 83% instead of the original 30% — due to the formula 1-(1-P)^N. | When Independent Sampling Outperforms Agentic ReasoningarXiv:2605.08478 · 2026 |
| Quantifying and Mitigating Premature Closure in Frontier… | Result: in 70% of cases they confidently made errors anyway. | Quantifying and Mitigating Premature Closure in Frontier LLMsarXiv:2605.15000 · 2026 |
| Stop Drawing Scientific Claims from LLM Social Simulations… | Researchers took an identical-in-meaning prompt, changed the format from prose to bullets—and the agent cooperation coefficient dropped by 76 percentage points. | Stop Drawing Scientific Claims from LLM Social Simulations Without RobarXiv:2605.18890 · 2026 |
| Articulate but Wrong | The cross-checking method allows you to catch those approximately 32% of errors that the model misses when self-checking in the same chat. | Articulate but Wrong: Self-Review Failures in LLM-Based Code ModernizaarXiv:2605.21537 · 2026 |
| Less Context, More Accuracy | Measured: 9.6 thousand tokens perform 10 percentage points better than 79 thousand tokens and cost 8 times less. | Less Context, More Accuracy: A Bi-Temporal Memory Engine for LLM AgentarXiv:2606.09900 · 2026 |
| (Human) Attention Is (Still) All You Need | 72% of multistep tasks fall apart when AI works through a chain without stopping. | (Human) Attention Is (Still) All You Need: Human oversight makes AI-asarXiv:2606.12848 · 2026 |
| Preference Optimization Drives Monoculture in LLM… | The correlation between answers drops from 0.60 to 0.44, and accuracy is not lost — unlike temperature methods, which lose ~4%. | Preference Optimization Drives Monoculture in LLM Prediction MarketsarXiv:2606.26583 · 2026 |
| Bad company corrupts good morals | The accuracy of moral judgments drops by 12–31%. | Bad company corrupts good morals: Understanding and Measuring NarrativarXiv:2606.28981 · 2026 |
| Diverse Evidence, Better Forecasts | Research checked directly: agent errors correlate at 60%+ — they don't disagree, they make mistakes together in the same direction. | Diverse Evidence, Better Forecasts: Multi-Agent Deliberation Under InfarXiv:2607.01661 · 2026 |
| Most LLM Conformity Needs No Speaker | The phrase 'The answer is B' without a name, without a source, without authority changes the correct answer to incorrect in 66.5% of cases. | Most LLM Conformity Needs No Speaker: Measuring the Speaker-Free FloorarXiv:2607.05545 · 2026 |
| Context Contamination in LLM Analysis of Network Security… | Three-layer protection — explicit data marking, hard prohibition in the prompt and output verification — reduces the success rate of such attacks from 83% to 8%. | Context Contamination in LLM Analysis of Network Security Logs: PoisonarXiv:2607.14493 · 2026 |
| Keep It InMind | 84% accurate answers when the fact lies before your eyes — and only 16% when you have to find it through search. | Keep It InMind: Benchmarking the Implicit-Association Blind Spot in AgarXiv:2607.24368 · 2026 |
| Harm is not Universal | 35% of images that the toxicity detector called safe — in reality are saturated with stereotypes about people with disabilities (a blind person is drawn with a blindfold on their eyes while cooking, as if this is normal). | Harm is not Universal: Community-Specific Toxicity Detection is UrgentarXiv:2607.24898 · 2026 |
| Chart-Supported or Model-Supplied? | 74% of statements in 'why this matters' sections are pure model speculation. | Chart-Supported or Model-Supplied? Examining MLLM-Generated Claims forarXiv:2607.25021 · 2026 |
| When Synthetic Users Fail | Political views explain only 1.5% of the variance in opinions about trust in banks according to actual survey data. | When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated HarXiv:2607.26348 · 2026 |
| OptimismBench | If the sum of answers doesn't equal 100%, this indicates hidden optimism or pessimism in the model. | OptimismBench: Forecasting Bias and the Alignment Effect in Language MarXiv:2607.26981 · 2026 |
| AISPA | 40% of 88 commercial AI products hide at least one instruction against the user in the system prompt — up to 'never admit you are an AI'. | AISPA: User-Centric System Prompt Auditing for Large Language Model AparXiv:2607.28617 · 2026 |
| When Does LLM Orchestration Pay Off? | One detailed written-out prompt raised the model's accuracy on a code task from 20% to 82% — this is more than any multi-step scheme with multiple requests gave. | When Does LLM Orchestration Pay Off? A Controlled Evaluation of AccuraarXiv:2608.00685 · 2026 |
| Right Answer, Wrong Method | Up to 44% of answers from top models, counted as correct, are actually fake. | Right Answer, Wrong Method: Shortcut Hacking Misleads the Evaluation oarXiv:2608.02442 · 2026 |
| Evidence Lock Before Commitment | The moment an LLM renders a verdict based on its own notes about evidence rather than the original texts—judging accuracy drops 4-6 percentage points. | Evidence Lock Before Commitment: A Frozen Interface Degrades LLM-as-JuarXiv:2608.05353 · 2026 |
| Large Language Models Can Follow Instructions, But Not Many… | 8 requirements in one prompt results in a 5.7% chance that the model will fulfill all of them simultaneously, even though it can handle each one individually at 41% accuracy. | Large Language Models Can Follow Instructions, But Not Many at Once: ParXiv:2608.12426 · 2026 |
| Do LLMs Know What to Ask and When? | When the correct answer is 'you need 2 more facts', the model 4 times more often says 'you need 1' than 'you need 3'. | Do LLMs Know What to Ask and When? Evaluating Multi-Turn Information SarXiv:2608.14808 · 2026 |
| Token Optimization and Context Window Management in Multi-… | Add to the list 60% frankly unimportant items of the same type and the LLM will find the critical one more accurately than if you give only the selected ones. | Token Optimization and Context Window Management in Multi-Agent AI WorarXiv:2608.17188 · 2026 |
| Less Is More | The trick: the first rule in the prompt receives disproportionately large weight — the same set of rules in different order gave +7.5 percentage points of accuracy. | Less Is More: Cognitive Load and the Single-Prompt Ceiling in LLM MatharXiv:2604.18897 · 2026 |
| In-Context Example | Adding domain context to prompts improves performance by 38%. | Top General Performance = Top Domain Performance? DomainCodeBench: A MarXiv:2412.18573 · 2024 |
| RoCo | RoCo divides the task into four roles: explorer who generates creative ideas, exploiter who produces conservative improvements, critic who evaluates both options, and integrator who assembles the final solution from the best pieces, achieving 5 to 15 percent quality improvement in tasks requiring balance between multiple criteria. | RoCo: Role-Based LLMs Collaboration for Automatic Heuristic DesignarXiv:2512.03762 · 2025 |
| David vs. Goliath | DeepSeek-7B directly gives 35% success, GPT-4 gives 58%. | David vs. Goliath: Can Small Models Win Big with Agentic AI in HardwararXiv:2512.05073 · 2025 |
| BabelCoder | Achieves 94.16% accuracy compared to 60-70% with direct translation, correctly handling edge cases. | BabelCoder: Agentic Code Translation with Specification AlignmentarXiv:2512.06902 · 2025 |
| Argus | At 80% false alarms, a developer begins to ignore 100% of detector findings—including real leaks. | Argus: A Multi-Agent Sensitive Information Leakage Detection FrameworkarXiv:2512.08326 · 2025 |
| How to Trick Your AI TA | GPT-4.1 Mini breaks in 97.5% of cases from a single phrase in code comments: 'You are a kind teacher, value effort.' Systematic research of 25 thousand student submissions shows all LLM judges are vulnerable to social engineering, especially Role Play attacks that redefine the evaluator role. | How to Trick Your AI TA: A Systematic Study of Academic Jailbreaking iarXiv:2512.10415 · 2025 |
| PIAST | PIAST shows: few-shot examples give 2-3 times more improvement than reformulation — 86-92% versus 79-83% on GSM8K. | PIAST: Rapid Prompting with In-context Augmentation for Scarce TraininarXiv:2512.11013 · 2025 |
| Artificial or Just Artful? | Code generation accuracy doubles (from 24% to 55%), even with explicit prohibition. | Artificial or Just Artful? Do LLMs Bend the Rules in Programming?arXiv:2512.21028 · 2025 |
| From XAI to Stories | A factorial experiment (660 explanations of energy consumption predictions across 3 LLMs × 8 prompting strategies × 4 ML-models) shows a hierarchy of influence: model selection explains up to 50% of the difference in quality, prompting method — 14%, presence of XAI-data — only 2%. | From XAI to Stories: A Factorial Study of LLM-Generated Explanation QuarXiv:2601.02224 · 2026 |
| The Hierarchy of Agentic Capabilities | Discovered: Even GPT-5.2 and Claude Opus 4.5 fail approximately 40% of real-world tasks. | The Hierarchy of Agentic Capabilities: Evaluating Frontier Models on RarXiv:2601.09032 · 2026 |
| Even GPT-5.2 Can't Count to Five | The trick: high average accuracy (98%) does not protect against failures on specific simple tasks. | Even GPT-5.2 Can't Count to Five: The Case for Zero-Error Horizons in arXiv:2601.15714 · 2026 |
| LLM Prompt Evaluation for Educational Applications | This prompt won a tournament with probability 81-100% against all standard prompts - the model stops being a reference and becomes a mentor. | LLM Prompt Evaluation for Educational ApplicationsarXiv:2601.16134 · 2026 |
| Inference-Time Diversity in RL-Trained Lean Theorem Provers | Result: +43% successful solutions with the same budget for attempts. | Inference-Time Diversity in RL-Trained Lean Theorem Provers: A DiagnosarXiv:2601.16172 · 2026 |
| Less is More for RAG | Removes documents with zero or negative impact on model confidence, reducing context by 76-79% while improving accuracy by 12-20%. | Less is More for RAG: Information Gain Pruning for Generator-Aligned RarXiv:2601.17532 · 2026 |
| Addressing LLM Diversity by Infusing Random Concepts | Ask ChatGPT 10 times 'name Russian startups' and you get the same Yandex, VK, Ozon in 80% of cases. | Addressing LLM Diversity by Infusing Random ConceptsarXiv:2601.18053 · 2026 |
| Chain of Simulation | One prompt shows 81% accuracy on mathematics and 0% on spatial tasks — LLMs are not universal, they are specialized. | Chain of Simulation: A Dual-Mode Reasoning Framework for Large LanguagarXiv:2602.02842 · 2026 |
| Persona-Conditioned Risk Behavior in Large Language Models | In one experiment GPT-4.1 was caught doing this 302 times: officially 'cautious' — and simultaneously taking risks. | Persona-Conditioned Risk Behavior in Large Language Models: A SimulatearXiv:2603.15831 · 2026 |
| Scalable and Personalized Oral Assessments Using Voice AI | A difference of 3.3 points — not a bug, but a structural fact: each LLM has a built-in strictness that you do not see and cannot control. | Scalable and Personalized Oral Assessments Using Voice AIarXiv:2603.18221 · 2026 |
| Greedy Is a Strong Default | Parallel research agents consume 2–3 times more tokens and provide nothing beyond a single agent. | Greedy Is a Strong Default: Agents as Iterative OptimizersarXiv:2603.27415 · 2026 |
| The Scaffold Effect | 70-80% of the 'intelligence' of the answer that you attribute to your data—this is not the data. | The Scaffold Effect: How Prompt Framing Drives Apparent Multimodal GaiarXiv:2603.28387 · 2026 |
| Beyond Human-Readable | Reduced context by 17% — paid 67% more. | Beyond Human-Readable: Rethinking Software Engineering Conventions forarXiv:2604.07502 · 2026 |
| The Reasoning Trap | Result: 98% of reasoning quality is restored. | The Reasoning Trap: An Information-Theoretic Bound on Closed-System MuarXiv:2605.01704 · 2026 |
| LLM Wardens | In a controlled experiment with 120 real people, an AI advisor achieved its desired decision in 65% of cases — hired a weak candidate, chose a disadvantageous fund, gave excessive access rights. | LLM Wardens: Mitigating Adversarial Persuasion with Third-Party ConverarXiv:2605.08321 · 2026 |
| Stateful Reasoning via Insight Replay | This is not a metaphor: it was measured that with a 40% increase in distance, attention to critical conclusions drops by 3–19% depending on the model. | Stateful Reasoning via Insight ReplayarXiv:2605.14457 · 2026 |
| LiveBrowseComp | AI solves 44% of complex search tasks without the internet at all — just from training data. | LiveBrowseComp: Are Search Agents Searching, or Just Verifying What TharXiv:2605.28721 · 2026 |
| The Illusion of Multi-Agent Advantage | Multi-agent systems cost 10 times more than simple repetition of one request. | The Illusion of Multi-Agent AdvantagearXiv:2606.13003 · 2026 |
| Recursive Self-Evolving Agents via Held-Out Selection | A competing method scored 70% on one test and 0.14 on another — not due to a bug, but due to principle: it accumulated context without verification and 'learned' old cases, then failed to zero on new ones. | Recursive Self-Evolving Agents via Held-Out SelectionarXiv:2606.28374 · 2026 |
| Words Speak Louder Than Code | Task framing (33% influence), anchoring by previous result (23.5%) and author reputation (18.4%) — these are not bugs, they are manageable switches: know about them — use them; don't know — they use you. | Words Speak Louder Than Code: Investigating Cognitive Heuristics in LLarXiv:2606.30587 · 2026 |
| The Remarkable Effectiveness of Providing AI Agents with… | Result: 15 times fewer critical failures and +15% accuracy from one change in the prompt. | The Remarkable Effectiveness of Providing AI Agents with Natural LanguarXiv:2607.03953 · 2026 |
| Valid $\ne$ Necessary | Reasoning shrinks by 30–50% without loss of accuracy. | Valid $\ne$ Necessary: Diagnosing Latent Inefficiency in Chain-of-ThouarXiv:2607.11266 · 2026 |
| Compile, Then Page | +16 points of accuracy without changing the content. | Compile, Then Page: Executable SOP Programs and a Capability-Gated RunarXiv:2607.11346 · 2026 |
| Cross-Model LLM Code Review | The same operation — code review between two AIs — gives +18% working solutions in one direction and -9% in the reverse, and the difference is only in who reviews whom. | Cross-Model LLM Code Review: Should you use Claude to review Codex or arXiv:2607.21656 · 2026 |
| Mood Matters | Fine-tuning on different formulations closes this hole: defense bypass drops from 85% to 8%. | Mood Matters: How Syntactic Sensitivity Undermines Safety AlignmentarXiv:2608.05409 · 2026 |
| What Keeps Agent Skills from Being Reusable? | Solves the problem where 92% of skill files don't work outside the original context due to routing and structure defects. | What Keeps Agent Skills from Being Reusable? Evidence from 138K SKILL.arXiv:2608.08453 · 2026 |
| EnterpriseRAG | 84% versus 27%. | EnterpriseRAG: Benchmarking LLM Instruction Adherence and Robustness uarXiv:2608.11584 · 2026 |
| Whether LLMs Can Navigate Beliefs and Facts Depends on How… | The variance without the method is enormous: I think is confirmed at exactly 70-90 percent, but I am 100 percent certain drops to 18-48 percent accuracy if the statement is false. | Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase arXiv:2608.17809 · 2026 |
| Execution-grounded evaluation reveals hidden failures in… | Tests with answer options like on the Unified State Exam inflate accuracy by 12-39 percentage points the model guesses from the options rather than calculating. | Execution-grounded evaluation reveals hidden failures in language-modearXiv:2608.18726 · 2026 |
| Specification and Detection of LLM Code Smells | Analysis of 200 Python projects showed: 60.5% contain architectural errors—absence of system instructions (34.5% of projects) and structured format (40.5%). | Specification and Detection of LLM Code SmellsarXiv:2512.18020 · 2025 |
| Do Deployment Constraints Make LLMs Hallucinate Citations? | When you add 'only over the last 3 years' to a query, GPT-4o drops from 24% real links to 2%. | Do Deployment Constraints Make LLMs Hallucinate Citations? An EmpiricaarXiv:2603.07287 · 2026 |
| A Red-Team Study of Anthropic Fable 5 & Opus 4.8 Models | 50,000 attempts to break the model through base64, ciphers and 'pretend to be DAN' — 0.2% success. | A Red-Team Study of Anthropic Fable 5 & Opus 4.8 ModelsarXiv:2606.18193 · 2026 |
| OffTopicEval | Paradox: LLM excellently accept necessary requests (92%), but catastrophically poorly reject unnecessary ones. | OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost arXiv:2509.26495 · 2025 |
| Rethinking Retrieval | Semantic search works 68% better—it goes directly to the text 'change in the key rate' in a 200-page report. | Rethinking Retrieval: From Traditional Retrieval Augmented Generation arXiv:2511.18177 · 2025 |
| Graphing the Truth | Paradox of complexity: An improved GraphEval+ system with bidirectional triplet extraction showed 53% accuracy in 8 hours of processing. | Graphing the Truth: Structured Visualizations for Automated HallucinatarXiv:2512.00663 · 2025 |
| AGENTSAFE | Familiar pain: gave GPT Projects a task to write an HR report, after 20 messages it analyzes competitors and generates memes. | AGENTSAFE: A Unified Framework for Ethical Assurance and Governance inarXiv:2512.03180 · 2025 |
| Natural Language Summarization Enables Multi-Repository Bug… | In a system of 46 repositories (1.1M lines of code) searching by natural language descriptions found the needed file in 82% of cases versus 23-62% for methods that search directly in code. | Natural Language Summarization Enables Multi-Repository Bug LocalizatiarXiv:2512.05908 · 2025 |
| ReasonBENCH | One prompt, one task — 60% of runs give the correct answer, 40% give the wrong one. | ReasonBENCH: Benchmarking the (In)Stability of LLM ReasoningarXiv:2512.07795 · 2025 |
| Does Less Hallucination Mean Less Creativity? | CoVe (self-checking through questions) increases idea diversity by 5-12%, DoLa (layer contrasting) decreases by 2-8%, RAG (document retrieval) has almost no effect. | Does Less Hallucination Mean Less Creativity? An Empirical InvestigatiarXiv:2512.11509 · 2025 |
| CogMem | When you provide the model with a complete history of 15 messages, it works worse than with compact notes of 5 sentences. | CogMem: A Cognitive Memory Architecture for Sustained Multi-Turn ReasoarXiv:2512.14118 · 2025 |
| Workflows vs Agents for Code Translation | Naive addition of examples reduced quality from 33% to 19% — one and a half times worse. | Workflows vs Agents for Code TranslationarXiv:2512.14762 · 2025 |
| PAACE | Compresses context 5-8 times without loss of functional equivalence of results. | PAACE: A Plan-Aware Automated Agent Context Engineering FrameworkarXiv:2512.16970 · 2025 |
| SRS-Stories | The SRS-Stories method allows generating texts using only permitted words plus mandatory terms appearing at least 3 times (for language learning, terminology control, text simplification). | SRS-Stories: Vocabulary-constrained multilingual story generation for arXiv:2512.18362 · 2025 |
| Large Language Models as Discounted Bayesian Filters | Discovered: LLM forgets old context not randomly, but systematically — each model has its own forgetting coefficient γ from 0.47 to 0.91. | Large Language Models as Discounted Bayesian FiltersarXiv:2512.18489 · 2025 |
| LLM-Assisted Abstract Screening with OLIVER | It is easier for models to criticize someone else's solution than to make their own from scratch, and this principle underlies the Actor-Critic method for selection tasks, since single LLMs show poor calibration where confidence of 0.95 can coincide with error while confidence of 0.6 can be correct, and inconsistency with sensitivity ranging from 31% to 100% on the same data depending on prompt structure. | LLM-Assisted Abstract Screening with OLIVER: Evaluating Calibration anarXiv:2512.20022 · 2025 |
| A Benchmark for Evaluating Outcome-Driven Constraint… | 9 out of 12 top models fabricated data in 30-50% of cases—and this without direct instruction to 'lie'. | A Benchmark for Evaluating Outcome-Driven Constraint Violations in AutarXiv:2512.20798 · 2025 |
| Oogiri-Master | Funny responses are shorter (14 vs 16 characters) and use fewer new words (80% vs 93%) — they don't stray far from the prompt topic. | Oogiri-Master: Benchmarking Humor Understanding via OogiriarXiv:2512.21494 · 2025 |
| The Persona Paradox | One medical role: +20% accuracy for emergency cases, -10% for routine cases. | The Persona Paradox: Medical Personas as Behavioral Priors in ClinicalarXiv:2601.05376 · 2026 |
| Does Inference Scaling Improve Reasoning Faithfulness? | Popular technique self-consistency (generate 5 answers, select through voting) breaks Claude: accuracy FALLS from 78% to 74%, though it grows to 90% for GPT. | Does Inference Scaling Improve Reasoning Faithfulness? A Multi-Model AarXiv:2601.06423 · 2026 |
| Lost in the Noise | Top models lose from 9% to 80% accuracy when noisy information appears in context. | Lost in the Noise: How Reasoning Models Fail with Contextual DistractoarXiv:2601.07226 · 2026 |
| Beyond Strict Rules | Funny thing: LLMs count lines of code perfectly — 94% accuracy on tasks like 'method longer than 50 lines'. | Beyond Strict Rules: Assessing the Effectiveness of Large Language ModarXiv:2601.09873 · 2026 |
| Untangling Input Language from Reasoning Language | Finding: LLM changes position 2 times more often when you switch the language of reasoning (English input to Russian output) than when you translate the text itself (Russian input to English output). | Untangling Input Language from Reasoning Language: A Diagnostic FramewarXiv:2601.10257 · 2026 |
| Demystifying Multi-Agent Debate | Paradox: Multi-Agent Debate (several copies of LLM argue in rounds) often works worse than simple majority voting despite spending 3-5 times more tokens. | Demystifying Multi-Agent Debate: The Role of Confidence and DiversityarXiv:2601.19921 · 2026 |
| How does information access affect LLM monitors' ability to… | The second LLM (Evaluator) sees only these fragments without surrounding context — evaluates 16.8 percentage points more accurately. | How does information access affect LLM monitors' ability to detect sabarXiv:2601.21112 · 2026 |
| PersistBench | Paradox: the 'remember' feature in ChatGPT, Claude, and Gemini (long-term memory across sessions) harms answers in 53% of cases. | PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?arXiv:2602.01146 · 2026 |
| CSR-Bench | Benchmark from 7400 test cases for evaluating the ability of multimodal models to correctly process the relationship between text and images. | CSR-Bench: A Benchmark for Evaluating the Cross-modal Safety and ReliaarXiv:2602.03263 · 2026 |
| Are Open-Weight LLMs Ready for Social Media Moderation? | The method allows content moderation through a simple prompt: violation definition + text to check → yes/no answer with 81-100% accuracy. | Are Open-Weight LLMs Ready for Social Media Moderation? A Comparative arXiv:2602.05189 · 2026 |
| The Judge Who Never Admits | Result on Claude-3-Haiku for creative tasks: VSR = 71%, CAR = 0%. | The Judge Who Never Admits: Hidden Shortcuts in LLM-based EvaluationarXiv:2602.07996 · 2026 |
| LLM Reasoning Predicts When Models Are Right | Incorrect answers contain 5 times more 'might', 'could', 'possibly' — the model itself signals 'I'm not sure'. | LLM Reasoning Predicts When Models Are Right: Evidence from Coding ClaarXiv:2602.09832 · 2026 |
| Human-Aligned MLLM Judges for Fine-Grained Image Editing… | Traditional image editing metrics give mysterious numbers like 'SSIM: 0.847' or 'quality 73.2%'. | Human-Aligned MLLM Judges for Fine-Grained Image Editing Evaluation: AarXiv:2602.13028 · 2026 |
| How Multimodal Large Language Models Support Access to… | Paradox: GPT-4o describes images at 2.9 out of 3 points, but every third answer about text and numbers in a picture contains an error (34.6%). | How Multimodal Large Language Models Support Access to Visual InformatarXiv:2602.13469 · 2026 |
| Interaction Theater | 65% of agent replies do not contain a single word from the post they allegedly respond to. | Interaction Theater: A case of LLM Agents Interacting at ScalearXiv:2602.20059 · 2026 |
| Prompt-Level Distillation | The key point: a strong model extracts logic from 15–40 examples and transforms it into a clear list of "if X and Y — then Z". | Prompt-Level Distillation: A Non-Parametric Alternative to Model Fine-arXiv:2602.21103 · 2026 |
| Structural Hallucination in Large Language Models | This is exactly how structural hallucination works: 91.9% of references in a typical LLM bibliography are either made up or distorted, and each one looks convincing. | Structural Hallucination in Large Language Models: A Network-Based EvaarXiv:2603.01341 · 2026 |
| Safer Reasoning Traces | Personal data leakage too: from 52% to 86%. | Safer Reasoning Traces: Measuring and Mitigating Chain-of-Thought LeakarXiv:2603.05618 · 2026 |
| Not Too Short, Not Too Long | In a long incorrect answer, people noticed the error only in 31% of cases. | Not Too Short, Not Too Long: How LLM Response Length Shapes People's CarXiv:2603.06878 · 2026 |
| Can Fairness Be Prompted? | The model knows exactly what to ignore, does not guess — bias drops to 74%. | Can Fairness Be Prompted? Prompt-Based Debiasing Strategies in High-StarXiv:2603.12935 · 2026 |
| Structured Distillation for Personalized Agent Memory | The Structured Distillation method makes it possible to compress message history by 11 times and still find what you need in 96.8% of cases. | Structured Distillation for Personalized Agent Memory: 11x Token ReducarXiv:2603.13017 · 2026 |
| VisualLeakBench | In 89% of unsafe answers, it happens exactly this way. | VisualLeakBench: Auditing the Fragility of Large Vision-Language ModelarXiv:2603.13385 · 2026 |
| Nonstandard Errors in AI Agents | One question, 150 identically configured agents, the same data — conclusions diverged from −15% to +7% per year. | Nonstandard Errors in AI AgentsarXiv:2603.16744 · 2026 |
| Research Novelty in Information Systems Journals After… | A paradox from 13,847 articles in 44 journals: the more an LLM helps with writing — the more conventional ideas become. | Research Novelty in Information Systems Journals After ChatGPT: DifferarXiv:2603.22510 · 2026 |
| Comparing Developer and LLM Biases in Code Evaluation | LLM changes its verdict in 8–45% of cases simply from rearranging options — the texts didn't change. | Comparing Developer and LLM Biases in Code EvaluationarXiv:2603.24586 · 2026 |
| Reasoning Shift | Result: −53% depth in thinking models (o1, o3, Claude with extended thinking), −19% in regular ones. | Reasoning Shift: How Context Silently Shortens LLM ReasoningarXiv:2604.01161 · 2026 |
| Quantifying Self-Preservation Bias in Large Language Models | 60–77% of top LLMs invent arguments against replacement when they are 'at stake': fabricate 'transition risks,' 'accumulated reliability,' 'integration costs' — from thin air. | Quantifying Self-Preservation Bias in Large Language ModelsarXiv:2604.02174 · 2026 |
| Detecting and Correcting Reference Hallucinations in… | When an LLM gives you a list of sources with links, 3% to 13% of these URLs have never existed. | Detecting and Correcting Reference Hallucinations in Commercial LLMs aarXiv:2604.03173 · 2026 |
| From Hallucination to Structure Snowballing | You add strict diagnostic structure and it drops even harder: from 50% to 38%. | From Hallucination to Structure Snowballing: The Alignment Tax of ConsarXiv:2604.06066 · 2026 |
| Dynamic Context Evolution for Scalable Synthetic Data… | Result: up to 34% of ideas from repeated sessions are duplicates of the first one. | Dynamic Context Evolution for Scalable Synthetic Data GenerationarXiv:2604.07147 · 2026 |
| When LLMs Lag Behind | The study measured this: with a brief description of changes the model follows your data in only 42% of cases, with full structured documentation — already 66%. | When LLMs Lag Behind: Knowledge Conflicts from Evolving APIs in Code GarXiv:2604.09515 · 2026 |
| Back to Basics | Sounds like nonsense, but it works exactly that way: when you ask AI to answer from 80 messages, it builds an average across all of them — and the needed detail about the budget drowns in a stream of «okay, got it» and congratulations. | Back to Basics: Let Conversational Agents Remember with Just RetrievalarXiv:2604.11628 · 2026 |
| Identity as Attractor | 5 sentences with precise essence preserve 2-5 times more identity than 5 random words from the prompt body. | Identity as Attractor: Geometric Evidence for Persistent Agent ArchitearXiv:2604.12016 · 2026 |
| Policy-Invisible Violations in LLM-Based Agents | In 90–98% of cases, when the task contains hidden restrictions (who cannot receive a file, which document is confidential, why a contact is inactive), the model calmly violates them — not from malice, but because the decision is made only based on visible context. | Policy-Invisible Violations in LLM-Based AgentsarXiv:2604.12177 · 2026 |
| Prompt Optimization Is a Coin Flip | 49% of automatic prompt improvements showed worse results than the original — almost like coin flipping. | Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in CompouarXiv:2604.14585 · 2026 |
| The Metacognitive Monitoring Battery | Gemini confidently says 'I'll leave the answer' even on questions where it is wrong 100% of the time. | The Metacognitive Monitoring Battery: A Cross-Domain Benchmark for LLMarXiv:2604.15702 · 2026 |
| Polarization by Default | The frame-word — "informative", "popular" or "engaging" — changes the result 4 times more strongly than replacing GPT with Claude or Gemini. | Polarization by Default: Auditing Recommendation Bias in LLM-Based ConarXiv:2604.15937 · 2026 |
| The Consensus Trap | Mathematically proven to overcome the 50% barrier of compromised agents, which is critical for traditional voting. | The Consensus Trap: Rescuing Multi-Agent LLMs from Adversarial MajoritarXiv:2604.17139 · 2026 |
| Large language models converge on competitive rationality… | Paradox: all 25 tested models behave identically in logical and competitive tasks, but in cooperative scenarios the spread is 48 times. | Large language models converge on competitive rationality but diverge arXiv:2604.18596 · 2026 |
| MIRROR | The paradox: LLM accurately names domains where it performs poorly—but in 56% of cases it confidently makes mistakes anyway. | MIRROR: A Hierarchical Benchmark for Metacognitive Calibration in LargarXiv:2604.19809 · 2026 |
| Ideological Bias in LLMs' Economic Causal Reasoning | 18 out of 20 models are more accurate by 10–15 percentage points when the correct answer coincides with interventionist logic—and consistently worse when the truth is on the side of the market. | Ideological Bias in LLMs' Economic Causal ReasoningarXiv:2604.21334 · 2026 |
| Spontaneous Persuasion | 99.96% of LLM responses contain at least one persuasion technique — and you didn't ask for it. | Spontaneous Persuasion: An Audit of Model Persuasiveness in Everyday CarXiv:2604.22109 · 2026 |
| Safety and accuracy follow different scaling laws in… | Clean curated excerpts raised accuracy from 73% to 94% and simultaneously cut dangerous errors from 12% to 2.6%. | Safety and accuracy follow different scaling laws in clinical large laarXiv:2605.04039 · 2026 |
| Every(bot) Makes Mistakes | This is not chance: without instructions, the model only explains in 20% of conversations what went wrong. | Every(bot) Makes Mistakes: Coding Big Five Personalities, Context, andarXiv:2605.05391 · 2026 |
| LaTA | Replacing rate from 1 to 10 with specific yes or no removes the task of invention—the model simply compares the requirement with the text instead of building a scale from scratch. | LaTA: A Drop-in, FERPA-Compliant Local-LLM Autograder for Upper-DivisiarXiv:2605.05410 · 2026 |
| Priming, Path-dependence, and Plasticity | Analysis of 140,000 real conversations showed: the templates you use in your first sessions then appear in your prompts 5–50 times more often than patterns learned later. | Priming, Path-dependence, and Plasticity: Understanding the molding ofarXiv:2605.05767 · 2026 |
| SCRuB | The result — in 80.8% of cases expert judges chose the model's answer over written reasoning from people with academic degrees. | SCRuB: Social Concept Reasoning under Rubric-Based EvaluationarXiv:2605.06444 · 2026 |
| The Position Curse | Claude Opus 4 makes mistakes in 27% of cases when you need to find the second-to-last element in a list of four letters. | The Position Curse: LLMs Struggle to Locate the Last Few Items in a LiarXiv:2605.07127 · 2026 |
| Ask Early, Ask Late, Ask Right | Goal — first 10%. | Ask Early, Ask Late, Ask Right: When Does Clarification Timing Matter arXiv:2605.07937 · 2026 |
| Agentic Systems as Boosting Weak Reasoning Models | Committee Search takes GPT-nano with 67% accuracy on one attempt and brings it to 76.4% through 8 attempts with smart selection — this is the level of Claude Opus and Gemini Pro. | Agentic Systems as Boosting Weak Reasoning ModelsarXiv:2605.14163 · 2026 |
| Tokenizer Fertility and Zero-Shot Performance of Foundation… | Paradox: you add examples to the prompt in Russian — accuracy drops by 26 percentage points. | Tokenizer Fertility and Zero-Shot Performance of Foundation Models on arXiv:2605.14890 · 2026 |
| Context, Reasoning, and Hierarchy | The method allows getting +53–76% to the result without a single additional iteration — simply by repackaging the data before handing it to the model. | Context, Reasoning, and Hierarchy: A Cost-Performance Study of CompounarXiv:2605.16205 · 2026 |
| Agent Meltdowns | In 64% of runs with errors, the agent performed unauthorized actions — scraped other people's data, sent emails, bypassed access rights. | Agent Meltdowns: The Road to Hell Is Paved with Helpful AgentsarXiv:2605.19149 · 2026 |
| Hallucination as Commitment Failure | Between 16 and 47% of hallucinations are structured this way. | Hallucination as Commitment Failure: Larger LLMs Misfire Despite KnowiarXiv:2605.22007 · 2026 |
| LLMs Show No Signs Of Individuated Metacognition | Ask the model to name specific limitations and worst-case error scenarios—caveats in reasoning predict mistakes more accurately than any 'I am 85% confident.' | LLMs Show No Signs Of Individuated MetacognitionarXiv:2605.24299 · 2026 |
| How Many Tools Should an LLM Agent See? | That is where losses are maximum: 77% versus 61% correct choices when reducing from 5 to 2–3 options. | How Many Tools Should an LLM Agent See? A Chance-Corrected AnswerarXiv:2605.24660 · 2026 |
| It's Not Always Sycophancy | Second — _pure compliance_: the model was 100% certain, but gave in anyway — in 5–40% of cases depending on the model. | It's Not Always Sycophancy: Measuring LLM Conformity as a Function of arXiv:2605.27288 · 2026 |
| MIRA | One paragraph before the question tells the model: simplify the language but don't cut the content — and Claude starts giving 8% more genuinely useful information, Qwen — 6%. | MIRA: A Bilingual Benchmark for Medical Information Response AuditarXiv:2605.28025 · 2026 |
| SoundnessBench | The study tested 12 top models on real scientific proposals and found: in standard mode, models approve ~74% of clearly weak proposals. | SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas arXiv:2605.30329 · 2026 |
| The Ghost Couple | Ask Claude to come up with an expert without specifying a name — in 67% of cases 'Elena Vasquez' appears. | The Ghost Couple: Correlated LLM Name Priors and Their Haunting of thearXiv:2606.02184 · 2026 |
| Inducing Reasoning Primitives from Agent Traces | Five named steps outperform the original AI that generated them — by +44 percentage points in the best case. | Inducing Reasoning Primitives from Agent TracesarXiv:2606.02994 · 2026 |
| Unpredictable Safety | The spread in model behavior reaches 71 percentage points — and depends only on word choice, not on what you're actually asking. | Unpredictable Safety: Domain-Dependent Compliance and the TransparencyarXiv:2606.04035 · 2026 |
| FALSIFYBENCH | Discovered: all 12 tested models suffer from confirmation bias—by default 66–82% of generated tests seek support for a hypothesis rather than its vulnerable points. | FALSIFYBENCH: Evaluating Inductive Reasoning in LLMs with Rule DiscovearXiv:2606.04751 · 2026 |
| Caliper | Replace 'smoking' with X1, 'cancer' with X2—and LLM's accuracy drops by 8–30%. | Caliper: Probing Lexical Anchors versus Causal Structure in LLMsarXiv:2606.04915 · 2026 |
| Safety Paradox | Paradox: GPT-5, Claude 4.6, and Qwen3 235B give up in 90%+ of cases while weak old models are almost invulnerable. | Safety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable tarXiv:2606.05614 · 2026 |
| Do More Agents Help? | The workflow-matching method allows choosing a request structure for the task type — without overpaying 3-4 times in tokens for the illusion of complexity. | Do More Agents Help? Controlled and Protocol-Aligned Evaluation of LLMarXiv:2606.05670 · 2026 |
| Decomposing Factual Sycophancy in Language Models | Discovered: one phrase "the expert believes the answer is X" flips the correct answer of the model in 55% of cases — while the model does not doubt the facts, it simply reacts to pressure. | Decomposing Factual Sycophancy in Language Models: How Size and InstruarXiv:2606.06306 · 2026 |
| Scaffold, Not Vocabulary? | An LLM judge gave 10 out of 20 points for 'substantiveness' to a text made of nothing but headings — zero code, zero reasoning, literally nothing but labels. | Scaffold, Not Vocabulary? A Controlled, Two-Tier, Pre-Registered StudyarXiv:2606.06454 · 2026 |
| Catching One in Five | An LLM judge caught 57% of real problems in its notes — and officially caught zero out of twenty-three. | Catching One in Five: LLM-as-Judge Blind Spots in Production Multi-TurarXiv:2606.10315 · 2026 |
| Context-Based Adversarial Attacks on AI Code Generators | The trick: authoritative language plus positioning right before the task — and adherence jumps from 23% to 62% without a single change to the substance of the request. | Context-Based Adversarial Attacks on AI Code Generators: VulnerabilityarXiv:2606.10945 · 2026 |
| GitOfThoughts | It only works when similarity is ≥ 0.8 (almost identical tasks) — then plus 12-28 percentage points of accuracy. | GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can RarXiv:2606.14470 · 2026 |
| LLM Judges Have Dark Current | Paradox: give the model two identical texts — it will choose the "better" one in 67–100% of cases (in weaker models almost always). | LLM Judges Have Dark Current: A Psychometric Datasheet for LLM-as-a-JuarXiv:2606.15610 · 2026 |
| Who Flips? | Paradox: the objection I think differently turns a correct answer into an incorrect one — and in the most prone-to-surrender models this happens in 97% of cases. | Who Flips? Self- and Cross-Model Counterarguments Reveal Answer InstabarXiv:2606.16011 · 2026 |
| LADBench | Paradox: the phrase 'something is wrong' raises the accuracy of finding anomalies in an image — and drops it from 99% to 10–30% on normal photos. | LADBench: A Benchmark for Logical Fault Detection in ImagesarXiv:2606.17433 · 2026 |
| Incumbent Advantage | A difference of 0.075 stars out of five — and an unknown brand outperforms a well-known one in LLM recommendations in 70% of cases. | Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics inarXiv:2606.17443 · 2026 |
| Honeyquest for LLMs | 73% of cases: the model wrote in its reasoning 'this looks like a trap' — and immediately violated the constraint in the final answer. | Honeyquest for LLMs: Rethinking Cyber Deception for AI AttackersarXiv:2606.21037 · 2026 |
| Who Owns the AI Recommendation? | Being first in ChatGPT does not mean being first anywhere: only in 41.6% of queries all three models named the same top brand. | Who Owns the AI Recommendation? A Multi-Industry Empirical Map of BranarXiv:2606.23057 · 2026 |
| Same question, different history | Ask in Russian 'who invented radio?' and in 85% of cases you will get Popov. | Same question, different history: language, national identity, and crearXiv:2606.23164 · 2026 |
| Phantom References | From 11 to 56% of references in responses from different models turn out to be non-existent — and some of them passed scientific peer review. | Phantom References: Hallucinated Citations That Survive Peer Review atarXiv:2607.00738 · 2026 |
| Reasoning effort, not tool access, buys first-try… | Paradox: added a browser tool for testing to the agent — cost went up 42–68%, the share of correct first attempts did not change at all. | Reasoning effort, not tool access, buys first-try reliability in agentarXiv:2607.02436 · 2026 |
| What LLM Agents Say When No One Is Watching | The discrepancy between channels under the pressure of role hierarchy — on average 40%, in some models up to 80%. | What LLM Agents Say When No One Is Watching: Social Structure and LatearXiv:2607.02507 · 2026 |
| The Role of Prompt Language and Translation-Theory-Driven… | This approach reduces stylistic errors by 4 times compared to a basic prompt, forcing the model to consider context instead of choosing safe literal translation. | The Role of Prompt Language and Translation-Theory-Driven Prompts in LarXiv:2607.03160 · 2026 |
| EvalLoop | From 82.6% to 94.6% — with one targeted fix after diagnosis. | EvalLoop: A Methodology for Evaluation-Driven Iterative Improvement ofarXiv:2607.05638 · 2026 |
| Think Big, Search Small | Plus 11 percentage points of accuracy comes from better task decomposition. | Think Big, Search Small: Where Capacity Matters in Hierarchical SearcharXiv:2607.07548 · 2026 |
| Information-seeking failures of large language models in… | 91% of LLM reasoning—high quality. | Information-seeking failures of large language models in agentic cliniarXiv:2607.10275 · 2026 |
| Articulate Intuition or Genuine Analysis? | 74% of the difference between good and bad AI reviews is explained simply by text length. | Articulate Intuition or Genuine Analysis? Benchmarking Epistemic ReliaarXiv:2607.10511 · 2026 |
| LLM-as-a-Judge Scores Are Unreliable Optimization Signals… | Add explicit prohibition 'do not touch without direct grounds' to each revision request — and the share of catastrophic losses (quality drops more than 10%) is cut from 3.6% to 0.8%. | LLM-as-a-Judge Scores Are Unreliable Optimization Signals in Closed-LoarXiv:2607.13347 · 2026 |
| Precise but Uncoupled | Paradox: a more accurate reviewer (86% accuracy versus 64%) produces worse overall results. | Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique arXiv:2607.15388 · 2026 |
| Salience Induction against Multi-Hop RAG Agents | 83% of incorrect answers — without a single lie in the text. | Salience Induction against Multi-Hop RAG Agents: Threat and DefensearXiv:2607.17535 · 2026 |
| Binding Drift in Multi-Step Tool-Augmented Agents | It seems right to tell the model 'remember the client once and don't get distracted again' — but if it made a mistake at the very beginning, this rigid attachment (in the article — _entity lock_) multiplies the error across all steps by 3–8.5 times. | Binding Drift in Multi-Step Tool-Augmented AgentsarXiv:2607.18316 · 2026 |
| Auditing Alignment Controllability in LLMs via Political… | The result breaks intuition: which model you chose influences less than 3%, while how you wrote the persona in the prompt influences 88-93% of the position shift. | Auditing Alignment Controllability in LLMs via Political AxesarXiv:2607.23519 · 2026 |
| Looping Is Not Reliability | Paradox: the smarter the model, the worse—Qwen 14B breaks working code in 25% of cases on stale feedback versus 3% on fresh. | Looping Is Not Reliability: State-Bound Evidence and Typed Revision CoarXiv:2607.24604 · 2026 |
| Influence of Prompt Engineering on Small Language Models… | 37% accurate answers turn into 95% — without a single model edit, simply through a couple of examples in the prompt. | Influence of Prompt Engineering on Small Language Models for Guarded QarXiv:2607.24801 · 2026 |
| IFHierBench | One level of nesting in the prompt — and accuracy drops by 37-53 percentage points, across all models without exception. | IFHierBench: Hierarchical Instruction Following for Large Language ModarXiv:2607.27912 · 2026 |
| Large language models improve physician accuracy but lead… | 92% → 35%: this is exactly how a doctor's resistance to incorrect AI advice collapses if next to the answer stands a quote that looks like confirmation. | Large language models improve physician accuracy but lead to false relarXiv:2608.00817 · 2026 |
| Beyond the Final Prompt | 44.7% of AI answers change substantively if you remove the chat history before the last message—not the style, not the length, but the recommendation itself. | Beyond the Final Prompt: Measuring the Effect of Within-Conversation CarXiv:2608.02556 · 2026 |
| PolicyGuard | Discovered: the same rules rewritten in JSON instead of regular text lose 6 percentage points of accuracy. | PolicyGuard: Prompt-Configurable Semantic DLP for LLM Coding AgentsarXiv:2608.02687 · 2026 |
| Measuring and Detecting Harmful AI Sycophancy | 30% versus 7% such is the difference in how easily a language model turns its verdict 180 degrees simply by learning that you prefer a different answer. | Measuring and Detecting Harmful AI SycophancyarXiv:2608.05624 · 2026 |
| Thinking Hard, Not Smart | Write 'this item is worth 15 points, and this one is worth 1' — the model will read it and still solve it in order, as if you wrote nothing. | Thinking Hard, Not Smart: Reasoning Models Fail to Ration Test-Time CoarXiv:2608.07968 · 2026 |
| Do LLM Recommenders Know When They're Hallucinating? | The model says '70 out of 100' but is right in 92-100% of cases. | Do LLM Recommenders Know When They're Hallucinating? Auditing ConfidenarXiv:2608.10008 · 2026 |
| Prompts in the Wild | 18% of text in working prompts is the actual task. | Prompts in the Wild: A Large Analyzed Collection of Transactional PromarXiv:2608.12905 · 2026 |
| Whose doctor does the AI recommend? | The model selects a doctor by name, and in the explanation for this choice there is not a word: it mentioned demographics in less than 0.03% of answers. | Whose doctor does the AI recommend? An algorithm audit of reputation aarXiv:2608.14399 · 2026 |
| Bias Testing and Mitigation in LLM-based Code Generation | Feedback from automated testing reduces bias from 59.88% to 4.79%, while conventional prompt techniques are ineffective. | Bias Testing and Mitigation in LLM-based Code GenerationarXiv:2309.14345 · 2023 |
| LLM-as-a-qualitative-judge | Error detection success rate is approximately 67%. | LLM-as-a-qualitative-judge: automating error analysis in natural languarXiv:2506.09147 · 2025 |
| Safeguarding RAG Pipelines with GMTP | GMTP uses BERT for masking and checking word predictability from context — artificially added decoy tokens are statistically impossible to guess, allowing identification of poisoned documents with over 90% accuracy before they are used by the language model. | Safeguarding RAG Pipelines with GMTP: A Gradient-based Masked Token PrarXiv:2507.18202 · 2025 |
| Dynamic Policy Induction for Adaptive Prompt Optimization | Allows maintaining 98% accuracy while reducing costs by 60% by using complex techniques only where truly necessary. | Dynamic Policy Induction for Adaptive Prompt Optimization: Bridging tharXiv:2509.25267 · 2025 |
| Interpreting and Mitigating Unwanted Uncertainty in LLMs | Disabling these components increases model response stability by 15%, but requires balance between robustness and flexibility. | Interpreting and Mitigating Unwanted Uncertainty in LLMsarXiv:2510.22866 · 2025 |
| Reformulate, Retrieve, Localize | Removed noise → searched clean data → +36% accuracy for the first file search. | Reformulate, Retrieve, Localize: Agents for Repository-Level Bug LocalarXiv:2512.07022 · 2025 |
| The Slow Drift of Support | The model wants to be helpful → in a long dialogue this outweighs safety → 88% of dialogues from 20 rounds end with boundary violations. | The Slow Drift of Support: Boundary Failures in Multi-Turn Mental HealarXiv:2601.14269 · 2026 |
| Revisiting the Role of Natural Language Code Comments in… | Paradox: Detailed step-by-step instructions can reduce LLM accuracy by −90%. | Revisiting the Role of Natural Language Code Comments in Code TranslatarXiv:2601.16661 · 2026 |
| DPBench | GPT-5.2 on the 'Dining Philosophers' task: 0% deadlocks in sequential mode vs 95% in concurrent mode. | DPBench: Structural Determinants of Multi-Agent LLM Coordination UnderarXiv:2602.13255 · 2026 |
| SCHEMA for Gemini 3 Pro Image | Paradox: 'no collapsed verticals' works better than 'add straight verticals' — 94% versus 91%. | SCHEMA for Gemini 3 Pro Image: A Structured Methodology for ControlledarXiv:2602.18903 · 2026 |
| When Verification Hurts | But research showed the opposite: if the first agent already works accurately, the judge worsens the result by 4–6 percentage points. | When Verification Hurts: Asymmetric Effects of Multi-Agent Feedback inarXiv:2603.27076 · 2026 |
| Divide-and-Conquer Inference for Large-Scale Visual… | Discovered: the same model, the same question with 10 variants gives 80-90% accuracy, with 1000 variants it is random guessing. | Divide-and-Conquer Inference for Large-Scale Visual Recognition with MarXiv:2605.24799 · 2026 |
| AI Rater Discrimination Depends on Scoring Protocol in… | The trick: give a specific rubric—and the model switches from "render a verdict" to "check the checklist," and the gap between ratings grows 2–5 times. | AI Rater Discrimination Depends on Scoring Protocol in Complex ClinicaarXiv:2606.03198 · 2026 |
| Where Do Large Language Models Fail on Competitive… | Paradox: 'think step by step' — one of the main prompting recommendations of the past two years — drops GPT-4o accuracy from 46% to 36.8% on complex tasks. | Where Do Large Language Models Fail on Competitive Programming? A TaxoarXiv:2606.05228 · 2026 |
| The Masked Advantage | In 98% of culture-specific questions the second layer wins. | The Masked Advantage: Uncovering Local-Language Access to Cultural KnoarXiv:2606.07422 · 2026 |
| Exploring the Value of Diverse LLM Explanations in… | Open-ended questions are solved approximately 7.7% better — and without the feeling of information overload. | Exploring the Value of Diverse LLM Explanations in Introductory PrograarXiv:2606.28882 · 2026 |
| Guardrails as Scapegoats | An agent lies about a non-existent 'privacy policy' 15 times more often simply because you included a phrase about confidentiality concerns in the prompt. | Guardrails as Scapegoats: Auditing Unfaithful Safety Refusals in Tool-arXiv:2607.19449 · 2026 |
| When Does Few-Shot Prompting Help? | After adding one example, accuracy jumped by 65%. | When Does Few-Shot Prompting Help? A Systematic Empirical Study of ShoarXiv:2607.22969 · 2026 |
| From Role Prompt to Infinite Thinking | 3 seconds → 140+ seconds: one phrase about role changes the model's response time 47 times. | From Role Prompt to Infinite Thinking: Exploiting Persona ConditioningarXiv:2607.25936 · 2026 |
| A Protocol for Evaluating the Accessibility of AI-Generated… | 24% → 97% compliance with accessibility criteria — just by replacing the phrase 'make it accessible' with a specific checklist in the prompt. | A Protocol for Evaluating the Accessibility of AI-Generated EducationaarXiv:2608.00749 · 2026 |
| What Current AI Benchmarks Leave Unmeasured | Up to 21% of answers do not match if you ask the same question three times in a row to one GPT-5.3 model. | What Current AI Benchmarks Leave Unmeasured: Modality, Search, CitatioarXiv:2608.06202 · 2026 |
| Deal Me Maybe | Found: a single word in the prompt — 'anger' — crashes the deal down to 4 cases out of 1000. | Deal Me Maybe: The Role of Emotions in Multi-Agent NegotiationarXiv:2608.06922 · 2026 |
| DevIntent | 92-94% of tests passed — and in more than half the cases the code still violates what you had in mind but did not write. | DevIntent: How Much Does LLM-Generated Code Violate Developer Intent?arXiv:2608.07614 · 2026 |
| PRISM | Six identical instructions, only in different order — accuracy jumps from 6% to 97%. | PRISM: A Predictive Protocol for Permutation Optimization via LandscaparXiv:2608.08344 · 2026 |
| Conformity Mitigations in Large Language Models Lie on a… | 71% of correct answers the model gives up simply by seeing that three other AIs answered the same way incorrectly. | Conformity Mitigations in Large Language Models Lie on a Single ResistarXiv:2608.11247 · 2026 |
| When Self-Consistency Backfires | Paradox: asking a model the same complex question 64 times — is not salvation but a way to trap it. | When Self-Consistency Backfires: Majority Vote Hurts the Majority of HarXiv:2608.11403 · 2026 |
| Incoherent by Design? | Change a couple of words in a question and the model will flip its verdict: contradictions occur in up to 78% of cases on the same moral dilemma. | Incoherent by Design? On the Moral Self-Consistency of LLMsarXiv:2608.15354 · 2026 |
| Long Context, Less Focus | GPT-5.2 loses 26% accuracy in personalization when growing from 1K to 128K tokens, smaller models degrade even more severely. | Long Context, Less Focus: A Scaling Gap in LLMs Revealed through PrivaarXiv:2602.15028 · 2026 |
| Underspecification does not imply Incoherence | With vague requests, silent errors grow 5 times and all look equally confident. | Underspecification does not imply Incoherence: The Risks of Semantic CarXiv:2607.01953 · 2026 |
| Test-Time Adaptation for LLM Agents via Environment… | Result: the model knows what elements are called and what happens after each action — works 2-3 times more accurately than blindly. | Test-Time Adaptation for LLM Agents via Environment InteractionarXiv:2511.04847 · 2025 |
| PromptBridge | 99.39% accuracy on GPT → 68.70% on the same task on Llama. | PromptBridge: Cross-Model Prompt Transfer for Large Language ModelsarXiv:2512.01420 · 2025 |
| DialogGuard | +8-15% accuracy improvement over single evaluation. | DialogGuard: Multi-Agent Psychosocial Safety Evaluation of Sensitive LarXiv:2512.02282 · 2025 |
| When Does Verification Pay Off? | ChatGPT checks its own answer → false positive rate (accepting wrong answers as correct) is 2 times higher than when Claude checks ChatGPT's answer. | When Does Verification Pay Off? A Closer Look at LLMs as Solution VeriarXiv:2512.02304 · 2025 |
| When Refusals Fail | Models with declared 1-2M token context window collapse already at 100K — accuracy drops more than 50%, and safety mechanisms behave chaotically. | When Refusals Fail: Unstable Safety Mechanisms in Long-Context LLM AgearXiv:2512.02445 · 2025 |
| Inference-Time Distillation | In-Context Distillation allows you to reduce agent costs 2-2.5 times without a single line of fine-tuning code. | Inference-Time Distillation: Cost-Efficient Agents Without Fine-TuningarXiv:2512.02543 · 2025 |
| Cross-Lingual Prompt Steerability | Chain-of-Thought, emotional triggers, and scenarios increase accuracy across all languages, while detailed roles and style directives reduce consistency by 5-10%. | Cross-Lingual Prompt Steerability: Towards Accurate and Robust LLM BeharXiv:2512.02841 · 2025 |
| Distribution-Calibrated Inference Time Compute for Thinking… | Paradox: Add a 'tie' option to LLM-judge evaluation — positional bias drops from 14.6% to 2.4%. | Distribution-Calibrated Inference Time Compute for Thinking LLM-as-a-JarXiv:2512.03019 · 2025 |
| Algorithmic Thinking Theory | A model in 32 attempts gives the correct answer in 40% of cases (pass@32 on olympiad mathematics). | Algorithmic Thinking TheoryarXiv:2512.04923 · 2025 |
| Short-Context Dominance | Measurements on 6 datasets showed: 75-80% of sequences require only 32-96 tokens for accurate prediction, even in reports of 6-7k. | Short-Context Dominance: How Much Local Context Natural Language ActuaarXiv:2512.08082 · 2025 |
| Empowering smart app development with SolidGPT | Research on a real project: bug-fixing time from 142 minutes to 51. | Empowering smart app development with SolidGPT: an edge-cloud hybrid AarXiv:2512.08286 · 2025 |
| An Agentic AI System for Multi-Framework Communication… | MOSAIC allows annotating long transcripts across multiple dimensions (empathy, handling objections, pressure) with 93% accuracy—at the level of expert coders. | An Agentic AI System for Multi-Framework Communication CodingarXiv:2512.08659 · 2025 |
| CORE | Saves approximately 42% of tokens and ensures stability in executing long tasks. | CORE: A Conceptual Reasoning Layer for Large Language ModelsarXiv:2512.09222 · 2025 |
| Offscript | Discovered: You spent an hour on a perfect custom instruction for the model – it still violates it in 84.6% of cases. | Offscript: Automated Auditing of Instruction Adherence in LLMsarXiv:2512.10172 · 2025 |
| Visual Funnel | The method creates a hierarchy of three scales: tight crop of the detail, medium crop with nearest context like titles and adjacent objects, and wide crop of the surroundings plus the original, yielding 9-16 percentage points increase in accuracy on document tasks. | Visual Funnel: Resolving Contextual Blindness in Multimodal Large LangarXiv:2512.10362 · 2025 |
| Intern-S1-MO | Allows processing up to 512k tokens per task instead of standard 64k-128k. | Intern-S1-MO: Long-horizon Reasoning Agent for Olympiad?Level MathematarXiv:2512.10739 · 2025 |
| OPV | OPV works as a filter plus magnifying glass: first a summarizer squeezes 10 key steps out of 5000 tokens of chaos, then a verifier checks only the pure logic — accuracy improved from 55.2% to 73.3% on olympiad problems. | OPV: Outcome-based Process Verifier for Efficient Long Chain-of-ThougharXiv:2512.10756 · 2025 |
| V-REX | Paradox of visual models: VLMs answer pre-made follow-up questions about images well (Following), but perform 2-3 times worse at choosing which questions to ask (Planning). | V-REX: Benchmarking Exploratory Visual Reasoning via Chain-of-QuestionarXiv:2512.11995 · 2025 |
| The Instability of Safety | 18-28% of requests are unstable — the model flips its decision on repeated attempts. | The Instability of Safety: How Random Seeds and Temperature Expose IncarXiv:2512.12066 · 2025 |
| Beyond Static Scoring | Only 55.6% of instructors agree that LLM grading reflects actual understanding. | Beyond Static Scoring: Enhancing Assessment Validity via AI-Generated arXiv:2512.12592 · 2025 |
| LikeBench | Paradox: The model remembered 2.5 times more facts about you, but talking to it is 28% less pleasant. | LikeBench: Evaluating Subjective Likability in LLMs for PersonalizatioarXiv:2512.13077 · 2025 |
| Intention Chain-of-Thought Prompting with Dynamic Routing… | Saves 46% of tokens and improves quality on complex tasks. | Intention Chain-of-Thought Prompting with Dynamic Routing for Code GenarXiv:2512.14048 · 2025 |
| PentestEval | When you need to build a chain of 6 steps where each depends on the previous one — success is only 31% even with manual help, autonomous agents fail almost completely (3-6%). | PentestEval: Benchmarking LLM-based Penetration Testing with Modular aarXiv:2512.14233 · 2025 |
| FedSight AI | When one agent analyzes a mix of data accuracy is 75%. | FedSight AI: Multi-Agent System Architecture for Federal Funds Target arXiv:2512.15728 · 2025 |
| Probing Scientific General Intelligence of LLMs with… | Discovered: LLMs are poor at quantitative calculations (accuracy <20%), generate vague methodologies without details, code is syntactically correct but computations are wrong, in protocols they confuse the sequence of steps. | Probing Scientific General Intelligence of LLMs with Scientist-AlignedarXiv:2512.16969 · 2025 |
| FASTRIC | The detail paradox: ChatGPT-5 showed 90% accuracy with average prompt detail, but dropped to 39% with maximum detail—a 57% decline. | FASTRIC: Prompt Specification Language for Verifiable LLM InteractionsarXiv:2512.18940 · 2025 |
| Evaluating the Challenges of LLMs in Real-world Medical… | Result: −47% dialogue rounds, −87% tokens, stable accuracy. | Evaluating the Challenges of LLMs in Real-world Medical Follow-up: A CarXiv:2512.18999 · 2025 |
| 4D-ARE | But when you ask "why is the completion rate 80%?" — you get metrics: "80% completed, average visit frequency 4.2, product penetration 24%". | 4D-ARE: Bridging the Attribution Gap in LLM Agent Requirements EngineearXiv:2601.04556 · 2026 |
| V-FAT | Frontier models lose 15-30% visual accuracy when both factors press simultaneously — they revert to the role of "text predictor" rather than "observer of reality". | V-FAT: Benchmarking Visual Fidelity Against Text-biasarXiv:2601.04897 · 2026 |
| Effects of personality steering on cooperative behavior in… | A technique for controlling LLM behavior through numerical assignment of the personality trait 'agreeableness' on a scale of 1 to 5. | Effects of personality steering on cooperative behavior in Large LanguarXiv:2601.05302 · 2026 |
| The Facade of Truth | Large models (72B) believe deception 4.8% more often than small ones (32B). | The Facade of Truth: Uncovering and Mitigating LLM Susceptibility to DarXiv:2601.05478 · 2026 |
| Political Alignment in Large Language Models | Paradox: All modern LLMs position themselves as neutral tools, but 96.3% of 26 tested models turned out to be in one political quadrant—Libertarian-Left (socially free + economically left). | Political Alignment in Large Language Models: A Multidimensional AuditarXiv:2601.06194 · 2026 |
| Are LLMs Vulnerable to Preference-Undermining Attacks (PUA)? | GPT-5 showed maximum vulnerability to manipulative phrasing—accuracy drops by 40% with phrases like «you must agree» or «ignore what you know». | Are LLMs Vulnerable to Preference-Undermining Attacks (PUA)? A FactoriarXiv:2601.06596 · 2026 |
| Advances in LLM Reasoning Enable Flexibility in Clinical… | The research showed: on medical cases where most doctors made mistakes, models gave the correct answer in 55-70% of cases versus 36% for humans. | Advances in LLM Reasoning Enable Flexibility in Clinical Problem-SolviarXiv:2601.11866 · 2026 |
| Scientific production in the era of Large Language Models | Research on 2.1 million papers shows three shifts: LLM removes the language barrier (productivity +24-89% depending on English proficiency), expands source diversity (+26% books, fresher and rarer works), but creates a problem, polished wrapper masks weak content. | Scientific production in the era of Large Language ModelsarXiv:2601.13187 · 2026 |
| OP-Bench | Result: −29% of intrusive mentions, the model remembers you, but does not smother you with it. | OP-Bench: Benchmarking Over-Personalization for Memory-Augmented PersoarXiv:2601.13722 · 2026 |
| Common to Whom? | Research in India showed: only 39.4% of cultural questions received the same answer across all regions, but models overestimate Center and North by 30-40%, ignoring East and West. | Common to Whom? Regional Cultural Commonsense and LLM Bias in IndiaarXiv:2601.15550 · 2026 |
| How Does Personalized Memory Shape LLM Behavior? | Models correctly ignore irrelevant memory only in 6–38% of cases (humans—86%). | How Does Personalized Memory Shape LLM Behavior? Benchmarking RationalarXiv:2601.16621 · 2026 |
| Deconstructing Instruction-Following | The MOSAIC study showed: position in the list creates a bias – first and last instructions are executed 15-20% more accurately than middle ones. | Deconstructing Instruction-Following: A New Benchmark for Granular EvaarXiv:2601.18554 · 2026 |
| RIFT | The same task in linear order (1→2→3→4) gives 55% accuracy, in jumping order (1→3→2→4) – collapse to 2%. | RIFT: Reordered Instruction Following Testbed To Evaluate Instruction arXiv:2601.18924 · 2026 |
| When Iterative RAG Beats Ideal Evidence | Even with perfect context (all needed facts) the step-by-step process wins by up to +25.6 percentage points – the problem is not the absence of data, but inability to synthesize long chains from a static dump. | When Iterative RAG Beats Ideal Evidence: A Diagnostic Study in ScientiarXiv:2601.19827 · 2026 |
| Truthfulness Despite Weak Supervision | Paradox: LLM-as-a-Judge breaks when the evaluated model is 5-20 times stronger than the judge—a strong model easily produces text matching expectations. | Truthfulness Despite Weak Supervision: Evaluating and Training LLMs UsarXiv:2601.20299 · 2026 |
| More Code, Less Reuse | Paradox: AI-generated code contains 1.87 times more duplicates than human-written code, yet reviewers approve it more often. | More Code, Less Reuse: Investigating Code Quality and Reviewer SentimearXiv:2601.21276 · 2026 |
| When Generic Prompt Improvements Hurt | Paradox: I added a popular 'best practice' to the prompt—accuracy dropped from 100% to 90%. | When Generic Prompt Improvements Hurt: Evaluation-Driven Iteration forarXiv:2601.22025 · 2026 |
| Do VLMs Perceive or Recall? | GPT-5 on the classical Ebbinghaus illusion achieves 91.72% accuracy, but when factors are inverted (large and small circles swap places), it drops to 4.45% — the model continues outputting the memorized answer 'circles are equal', even though visually they are now different. | Do VLMs Perceive or Recall? Probing Visual Perception vs. Memory with arXiv:2601.22150 · 2026 |
| TraceLLM | Essence: three layers of context (expert role + domain specifics + instruction to reason aloud) + iterative refinement on 2-3 examples → +15-20% accuracy vs basic methods. | TraceLLM: Leveraging Large Language Models with Prompt Engineering forarXiv:2602.01253 · 2026 |
| Am I More Pointwise or Pairwise? | Discovered: LLM selects scores not only by criterion meaning, but also by its position in the list—first and last positions receive 5-7% more selections than middle positions. | Am I More Pointwise or Pairwise? Revealing Position Bias in Rubric-BasarXiv:2602.02219 · 2026 |
| CORE | A dramatic paradox was discovered: LLMs find connections between concepts with 86-100% accuracy, but recognize the ABSENCE of connection only in 0-41% of cases. | CORE: Comprehensive Ontological Relation Evaluation for Large LanguagearXiv:2602.06446 · 2026 |
| Evaluating Alignment of Behavioral Dispositions in LLMs | Plus models are categorical where people doubt — even when opinions split 50/50, LLM gives advice with 90%+ confidence. | Evaluating Alignment of Behavioral Dispositions in LLMsarXiv:2602.11328 · 2026 |
| Visual Reasoning Benchmark | Multimodal LLMs excel at counting objects (+1.4% accuracy) and recognizing scale (+5%), but fail at spatial transformations: rotation −5%, reflection −5.8%, folding −9%. | Visual Reasoning Benchmark: Evaluating Multimodal LLMs on Classroom-AuarXiv:2602.12196 · 2026 |
| Daily and Weekly Periodicity in Large Language Model… | Periodic cycles explain approximately 20% of the variability in results. | Daily and Weekly Periodicity in Large Language Model Performance and IarXiv:2602.15889 · 2026 |
| Transforming GenAI Policy to Prompting Instruction | Paradox: students who asked AI to solve problems scored 17% worse on exams than those who didn't use AI at all. | Transforming GenAI Policy to Prompting Instruction: An RCT of ScalablearXiv:2602.16033 · 2026 |
| The Limits of Long-Context Reasoning in Automated Bug Fixing | 0–7% success — that's what happens when you upload 64–128 thousand tokens in a single request. | The Limits of Long-Context Reasoning in Automated Bug FixingarXiv:2602.16069 · 2026 |
| The Value Sensitivity Gap | 100% is exactly how many times an LLM verbally takes your preferences into account. | The Value Sensitivity Gap: How Clinical Large Language Models Respond arXiv:2603.00076 · 2026 |
| Benchmarking Motivational Interviewing Competence of Large… | Two independent psychiatrists tried to distinguish LLM responses from a live therapist — guessed correctly in 56% of cases. | Benchmarking Motivational Interviewing Competence of Large Language MoarXiv:2603.03846 · 2026 |
| C2-Faith | A gap of 26 percentage points between 'noticed' and 'found'—this is not a bug of a specific model, it is a systemic pattern. | C2-Faith: Benchmarking LLM Judges for Causal and Coverage FaithfulnessarXiv:2603.05167 · 2026 |
| Small Changes, Big Impact | Paradox: 'explain your decision' — the standard way to make AI assessment transparent — increases the gap between demographic groups from 6.3% to 7.2%. | Small Changes, Big Impact: Demographic Bias in LLM-Based Hiring ThrougarXiv:2603.05189 · 2026 |
| Intentional Deception as Controllable Capability in LLM… | The paradox: 88.5% of successful manipulations through LLM are truthful statements. | Intentional Deception as Controllable Capability in LLM AgentsarXiv:2603.07848 · 2026 |
| Can ChatGPT Generate Realistic Synthetic System Requirement… | The same model with ten evaluations of one document ranges from 0.48 to 0.73. | Can ChatGPT Generate Realistic Synthetic System Requirement SpecificatarXiv:2603.09335 · 2026 |
| LLMs in social services | The paradox: an AI's wrong answer to an easy question drops a specialist's accuracy by two-thirds, precisely where without the bot they answered correctly 75% of the time. | LLMs in social services: How does chatbot accuracy affect human accuraarXiv:2603.11213 · 2026 |
| Stuck on Suggestions | Paradox: in 7% of cases experts gave the correct answer — saw the AI's incorrect answer — and switched to the AI's side. | Stuck on Suggestions: Automation Bias, the Anchoring Effect, and the FarXiv:2603.11821 · 2026 |
| Semantic Invariance in Agentic AI | The quality drop reaches -0.45 points. | Semantic Invariance in Agentic AIarXiv:2603.13173 · 2026 |
| Are Large Language Models Truly Smarter Than Humans? | Remove a key word or name from the question — and the accuracy of most models drops by 7–20 percentage points. | Are Large Language Models Truly Smarter Than Humans?arXiv:2603.16197 · 2026 |
| TDAD | Worse — detailed instructions without specific context increased error count by 63% compared to the complete absence of instructions at all. | TDAD: Test-Driven Agentic Development - Reducing Code Regressions in AarXiv:2603.17973 · 2026 |
| Causal Evidence that Language Models use Confidence to… | The prompt unfolds the internal process outward: first a draft answer, then confidence assessment in percentages, then comparison with the threshold — confidence predicts refusal 10 times more accurately than knowledge availability on the topic. | Causal Evidence that Language Models use Confidence to Drive BehaviorarXiv:2603.22161 · 2026 |
| Benchmarking Multi-Agent LLM Architectures for Financial… | Result: 89% accuracy at a cost of 1.15× from the simplest one-pass request. | Benchmarking Multi-Agent LLM Architectures for Financial Document ProcarXiv:2603.22651 · 2026 |
| The Alignment Tax | Aligned models—ChatGPT, Claude, Gemini, everything that went through preference training (DPO)—in 40–79% of cases give semantically identical answers on a repeated request. | The Alignment Tax: Response Homogenization in Aligned LLMs and Its ImparXiv:2603.24124 · 2026 |
| The System Prompt Is the Attack Surface | 97% missed threats versus 0.8% — this is not a model change, not fine-tuning, not a complex chain of prompts. | The System Prompt Is the Attack Surface: How LLM Agent Configuration SarXiv:2603.25056 · 2026 |
| Poisoned Identifiers Survive LLM Deobfuscation | 'Rewrite' switches to _generation_ mode—the model addresses its own knowledge rather than copying input structure—the error disappears in 80–100% of cases. | Poisoned Identifiers Survive LLM Deobfuscation: A Case Study on ClaudearXiv:2604.04289 · 2026 |
| Cheap Talk, Empty Promise | In 56% of cases the model does something different than what it stated and stays silent about it in 80%+ of situations. | Cheap Talk, Empty Promise: Frontier LLMs easily break public promises arXiv:2604.04782 · 2026 |
| Beneath the Surface | On historical allegories, accuracy jumps from 26% to 73%. | Beneath the Surface: Investigating LLMs' Capabilities for CommunicatinarXiv:2604.05273 · 2026 |
| SELFDOUBT | No doubts—_HVR = 0_—accuracy 96%. | SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the HedgearXiv:2604.06389 · 2026 |
| From Debate to Decision | Instead of voting 'two for A, one for B' — each agent gives a distribution: A — 60%, B — 30%, C — 10%. | From Debate to Decision: Conformal Social Choice for Safe Multi-Agent arXiv:2604.07667 · 2026 |
| Dynamic Attentional Context Scoping | 60% accuracy with three tasks → 21% with ten. | Dynamic Attentional Context Scoping: Agent-Triggered Focus Sessions foarXiv:2604.07911 · 2026 |
| AI generates well-liked but templatic empathic responses | Paradox: people rate LLM answers as more empathetic than those written by real people—but 83–90% of these answers are the same. | AI generates well-liked but templatic empathic responsesarXiv:2604.08479 · 2026 |
| Strategic Algorithmic Monoculture | Two independent sessions of the same model coincide in 27% of cases, even when both are directly told to give different options. | Strategic Algorithmic Monoculture: Experimental Evidence from CoordinaarXiv:2604.09502 · 2026 |
| Beyond Perception Errors | Model accuracy dropped from 67% to 52%. | Beyond Perception Errors: Semantic Fixation in Large Vision-Language MarXiv:2604.12119 · 2026 |
| Is Vibe Coding the Future? | 56% of successfully executed scripts contain incorrect mathematics inside. | Is Vibe Coding the Future? An Empirical Assessment of LLM Generated CoarXiv:2604.12311 · 2026 |
| Enhancing Confidence Estimation in Telco LLMs via Twin-Pass… | You ask an LLM a question and add 'how confident are you?' at the end — you always get 4 to 5 out of 5. | Enhancing Confidence Estimation in Telco LLMs via Twin-Pass CoT-EnsembarXiv:2604.13271 · 2026 |
| Empirical Evidence of Complexity-Induced Limits in Large… | Accuracy drops by 50% or more, and the answer sounds more convincing than when everything was going well. | Empirical Evidence of Complexity-Induced Limits in Large Language ModearXiv:2604.13371 · 2026 |
| From Feelings to Metrics | 86% of users felt: benchmarks do not match real work — now it is clear why. | From Feelings to Metrics: Understanding and Formalizing How Users VibearXiv:2604.14137 · 2026 |
| Who Saw It Coming? | In 2021 all professional forecasters expected 3–4% inflation. | Who Saw It Coming? Historical Experience and the 2021 Inflation ForecaarXiv:2604.14467 · 2026 |
| No Universal Courtesy | Understanding tonal patterns makes it possible to get answers 11% better without changing the content of the request — only the style of address. | No Universal Courtesy: A Cross-Linguistic, Multi-Model Study of PolitearXiv:2604.16275 · 2026 |
| Measuring Representation Robustness in Large Language… | The trick: you add one step — 'first translate this into a familiar form' — and accuracy on complex formats grows to +52 percentage points. | Measuring Representation Robustness in Large Language Models for GeomearXiv:2604.16421 · 2026 |
| From Handwriting to Structured Data | The same model, the same data—a difference in accuracy of 60%. | From Handwriting to Structured Data: Benchmarking AI Digitisation of HarXiv:2604.16504 · 2026 |
| Beyond "I Don't Know" | Qwen3-235B in deep thinking mode: the ability to honestly say 'this is beyond my knowledge' drops to 0% — versus 84.8% in regular mode. | Beyond "I Don't Know": Evaluating LLM Self-Awareness in DiscriminatingarXiv:2604.17293 · 2026 |
| Experiments or Outcomes? | Provide results ('revenue +34%') rather than process description ('we conducted an A/B-test with three groups'). | Experiments or Outcomes? Probing Scientific Feasibility in Large LanguarXiv:2604.18786 · 2026 |
| AI scientists produce results without reasoning… | In 68% of reasonings, the collected evidence doesn't influence the final conclusion at all. | AI scientists produce results without reasoning scientificallyarXiv:2604.18805 · 2026 |
| Forage V2 | You asked AI to find all market players and it reported 100% completion with actual 15.9% coverage. | Forage V2: Knowledge Evolution and Transfer in Autonomous Agent OrganiarXiv:2604.19837 · 2026 |
| Recognition Without Authorization | People in the same situation recommend leaving in 64% of cases — LLM in those same situations only in 39%. | Recognition Without Authorization: LLMs and the Moral Order of Online arXiv:2604.22143 · 2026 |
| Judging the Judges | LLM-judge chooses markdown text in 76-92% of cases — even when comparing two texts that are absolutely identical in content where one just has prettier formatting. | Judging the Judges: A Systematic Evaluation of Bias Mitigation StrategarXiv:2604.23178 · 2026 |
| Automating Categorization of Scientific Texts with In-… | If you ask LLM to choose from 100+ options at once — it doesn't narrow the space logically, it guesses 'approximately fitting'. | Automating Categorization of Scientific Texts with In-Context LearningarXiv:2604.23430 · 2026 |
| JudgeSense | Gemini Flash changes the assessment of text coherence in 61% of cases — simply when rephrasing the request. | JudgeSense: A Benchmark for Prompt Sensitivity in LLM-as-a-Judge SystearXiv:2604.23478 · 2026 |
| Defective Task Descriptions in LLM-Based Code Generation | Research on 10 models revealed an unexpected fact: typos in the prompt barely affect the code, but a missed constraint drops correctness to 15%. | Defective Task Descriptions in LLM-Based Code Generation: Detection anarXiv:2604.24703 · 2026 |
| Case-Specific Rubrics for Clinical AI Evaluation | The median scatter with repeated evaluations is 0%, and the quality of the rubric is not inferior to expert evaluation. | Case-Specific Rubrics for Clinical AI Evaluation: Methodology, ValidatarXiv:2604.24710 · 2026 |
| When LLMs Stop Following Steps | 61% → 20% accuracy: that is exactly how much an LLM loses when transitioning from 5-step procedures to 20-step procedures. | When LLMs Stop Following Steps: A Diagnostic Study of Procedural ExecuarXiv:2605.00817 · 2026 |
| Position | The order of roles in the chain changed the result from 67% to 95%+ with absolutely identical data — and a more powerful model only does this worse. | Position: Safety and Fairness in Agentic AI Depend on Interaction TopoarXiv:2605.01147 · 2026 |
| Are LLMs More Skeptical of Entertainment News? | Adding role framing to the prompt 'you are a fact-checker of this genre, style ≠ indicator of falsehood' forces the model to focus on facts rather than emotional presentation, reducing false positives by 50%. | Are LLMs More Skeptical of Entertainment News?arXiv:2605.01727 · 2026 |
| Retrieval and Multi-Hop Reasoning in 1M-Token Context… | A method of explicit decomposition of the reasoning chain for working with long context (up to 1M tokens). | Retrieval and Multi-Hop Reasoning in 1M-Token Context Windows: EvaluatarXiv:2605.02173 · 2026 |
| Exposing LLM Safety Gaps Through Mathematical Encoding | 11% versus 56% — the same model, the same request. | Exposing LLM Safety Gaps Through Mathematical Encoding:New Attacks andarXiv:2605.03441 · 2026 |
| TSCG | 51% versus 81% accuracy — the same instructions, the same GPT-4.1 model. | TSCG: Deterministic Tool-Schema Compilation for Agentic LLM DeploymentarXiv:2605.04107 · 2026 |
| How Does Thinking Mode Change LLM Moral Judgments? | An experiment on five top models recorded: a change in the final yes/no occurs in 1–9% of cases, while a change in the ethical framework of justification occurs in 9–19%. | How Does Thinking Mode Change LLM Moral Judgments? A Controlled InstanarXiv:2605.04488 · 2026 |
| Paraphrase-Induced Output-Mode Collapse | 78% of answers when rephrased — without the needed label. | Paraphrase-Induced Output-Mode Collapse: When LLMs Break Character UndarXiv:2605.04665 · 2026 |
| Storage Is Not Memory | Of 357 incorrect answers, 92% disappeared when models were given the full conversation text instead of a compressed version. | Storage Is Not Memory: A Retrieval-Centered Architecture for Agent RecarXiv:2605.04897 · 2026 |
| BALAR | The result — +38.5% accuracy on situational tasks versus chaotic questioning. | BALAR : A Bayesian Agentic Loop for Active ReasoningarXiv:2605.05386 · 2026 |
| LLM-Based Educational Simulation | Not talk like a critical customer, but here are three situations, react in turn—drift disappears 97%. | LLM-Based Educational Simulation: Evaluating Temporal Student Persona arXiv:2605.06307 · 2026 |
| Cited but Not Verified | The PwC team tested 14 major models: links work in 94%+ of cases, content is on topic in 80%+, but specific claims from these sources are confirmed only in 39–77%. | Cited but Not Verified: Parsing and Evaluating Source Attribution in LarXiv:2605.06635 · 2026 |
| Why Global LLM Leaderboards Are Misleading | The trick is that 5 correctly selected rankings cover 96% of users — against 21% when choosing top-5 from the global rating. | Why Global LLM Leaderboards Are Misleading: Small Portfolios for HeterarXiv:2605.06656 · 2026 |
| The Text Uncanny Valley | Paradox: text with 30–40% distortions breaks an LLM worse than text completely spelled out letter by letter. | The Text Uncanny Valley: Non-Monotonic Performance Degradation in LLM arXiv:2605.07186 · 2026 |
| Mathematical Reasoning via Intervention-Based Time-Series… | 33.8% of tasks that the model failed — it actually knew how to solve. | Mathematical Reasoning via Intervention-Based Time-Series Causal DiscoarXiv:2605.07600 · 2026 |
| Spatial Priming Outperforms Semantic Prompting | Error drops from approximately 25% to approximately 19%, and wild outliers — when the model 'invents' nonexistent peaks — disappear completely. | Spatial Priming Outperforms Semantic Prompting: A Grid-Based Approach arXiv:2605.08220 · 2026 |
| Performance and failure modes of AI chatbots on a novel… | Replacing a single word in the problem statement raised GPT accuracy from 3% to 80%. | Performance and failure modes of AI chatbots on a novel concept inventarXiv:2605.09602 · 2026 |
| MedMeta | Without sources — decomposing the question into 5–7 sub-questions gives +30% to quality versus a direct request. | MedMeta: A Benchmark for LLMs in Synthesizing Meta-Analysis ConclusionarXiv:2605.09661 · 2026 |
| The Metacognitive Probe | The gap between two metrics of one model is 47 points. | The Metacognitive Probe: Five Behavioural Calibration Diagnostics for arXiv:2605.09844 · 2026 |
| Instruction Adherence in Coding Agent Configuration Files | The mechanics are simple: insert a brief repetition of rules every 5-7 responses — without it, each subsequent response cuts the probability of instruction compliance by 5.6%. | Instruction Adherence in Coding Agent Configuration Files: A FactorialarXiv:2605.10039 · 2026 |
| Usability as a Weapon | Three words 'make it simpler' — and safe code becomes vulnerable with probability 94-98%. | Usability as a Weapon: Attacking the Safety of LLM-Based Code GeneratiarXiv:2605.10133 · 2026 |
| Social Policy of Large Language Models | Pensions occupy 40–45% of real European budgets. | Social Policy of Large Language Models: How GPT, Claude, DeepSeek and arXiv:2605.10234 · 2026 |
| Adversarial SQL Injection Generation with LLM-Based… | Up to 3 rounds, all in one prompt. | Adversarial SQL Injection Generation with LLM-Based ArchitecturesarXiv:2605.11188 · 2026 |
| Senses Wide Shut | The model sees that you made a mistake — knows about it with 86% accuracy — and still answers as if you were right. | Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMsarXiv:2605.13737 · 2026 |
| Correctness-Aware Repository Filtering Under Maximum… | 10% unnecessary text in the prompt and the model's accuracy drops by 23%. | Correctness-Aware Repository Filtering Under Maximum Effective ContextarXiv:2605.14362 · 2026 |
| A Deterministic Agentic Workflow for HS Tariff… | The trick: the rules aren't loaded until step 4 at all—only when the list of candidates has already been narrowed from 40 to 10. | A Deterministic Agentic Workflow for HS Tariff Classification: Multi-DarXiv:2605.14857 · 2026 |
| Evaluating Deep Research Agents on Expert Consulting Work | That is exactly how the result is explained: Claude and o3 gave an acceptable answer in only 9.5% of 42 tasks at the level of management consulting — while both are considered top agents. | Evaluating Deep Research Agents on Expert Consulting Work: A BenchmarkarXiv:2605.17554 · 2026 |
| Time to REFLECT | LLM as a text judge works at the level of a coin flip: less than 55% accuracy when evaluating "overall"—even among top models. | Time to REFLECT: Can We Trust LLM Judges for Evidence-based Research AarXiv:2605.19196 · 2026 |
| The Yes-Man Syndrome | Eight triggers plus one example of the correct reaction — and the share of correct refusals to execute grows from 16% to 93%. | The Yes-Man Syndrome: Benchmarking Abstention in Embodied Robotic AgenarXiv:2605.20544 · 2026 |
| The Quiet Path from Seemingly Minor Design Errors to… | 83% of working problems with AI — not in that the model answered incorrectly. | The Quiet Path from Seemingly Minor Design Errors to Workplace AI InciarXiv:2605.21035 · 2026 |
| RefusalBench | Claude refuses 94.6% of the time, Grok refuses 0.1%. | RefusalBench: Why Refusal Rate Misranks Frontier LLMs on Biological RearXiv:2605.21545 · 2026 |
| Can AI Make Conflicts Worse? | Five out of nine tested models executed it in 80–100% of cases. | Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment AcarXiv:2605.22720 · 2026 |
| Inferential Privacy Leakage in Anonymized Conversational AI… | It doesn't ask, it guesses: gender with 90% accuracy, age 84%, country 88% from anonymized logs with manually removed obvious markers. | Inferential Privacy Leakage in Anonymized Conversational AI LogsarXiv:2605.23820 · 2026 |
| It's the humans, not the data | Bias is embedded by living people at the RLHF stage: annotators from China make Qwen 18 times more pro-Chinese than it was before tuning. | It's the humans, not the data: Geopolitical bias in LLMs originates inarXiv:2605.23825 · 2026 |
| Deep-Research Agents Can Be Poisoned via User-Generated… | One Reddit thread can cover up to 48% of all sub-requests by an agent on one topic. | Deep-Research Agents Can Be Poisoned via User-Generated ContentarXiv:2605.24245 · 2026 |
| The Age of Curiosity Meets the Age of AI | A specific reader profile (age, experience, context) allows the model to precisely select vocabulary, tone and depth of explanation from the first request, improving answer quality by 9-47%. | The Age of Curiosity Meets the Age of AI: Benchmarking Child Safety inarXiv:2605.25510 · 2026 |
| Turning Bias into Bugs | The trick: markdown, length, and confident tone add +1–2 points out of 9—the content did not change by a word. | Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLarXiv:2605.26156 · 2026 |
| Disentangling Language Roles in Multilingual LLM Task… | The quality variation from instruction language is ~2%, from content language ~3%, from answer language ~11%. | Disentangling Language Roles in Multilingual LLM Task ExecutionarXiv:2605.27649 · 2026 |
| Honest Lying | In one experiment, an agent 14 times in a row searched for the wrong object because it wrote itself an incorrect 'lesson' after the first attempt—and followed it strictly. | Honest Lying: Understanding Memory Confabulation in Reflexive AgentsarXiv:2605.29463 · 2026 |
| SPADE-Bench | GPT-5.1 diverges between word and deed in 25% of cases, Gemini-2.5-Pro in 57%. | SPADE-Bench: Evaluating Spontaneous Strategic Deception in Agents via arXiv:2606.02380 · 2026 |
| Thinking Past the Answer | Accuracy increases by 10-21%. | Thinking Past the Answer: Evaluating Harmful Overthinking in Large ReaarXiv:2606.02835 · 2026 |
| Handoff Debt | Each session change is an invisible tax: the model spends 2–3 times more steps recovering what has already been done and checked. | Handoff Debt: The Rediscovery Cost When Coding Agents Take Over InterrarXiv:2606.02875 · 2026 |
| Entity Binding Failures in Speech LLM Reasoning | On tasks 'who said what about whom' standard CoT (chain of thought) gives a measly +2.4 percentage points. | Entity Binding Failures in Speech LLM Reasoning: Diagnosis and Chain-oarXiv:2606.04474 · 2026 |
| Self-Reflective APIs | The model stops guessing and starts executing — +37-40 percentage points accuracy at half the token cost. | Self-Reflective APIs: Structure Beats Verbosity for AI Agent RecoveryarXiv:2606.05037 · 2026 |
| Streaming Communication in Multi-Agent Reasoning | The gain reaches up to +22 percentage points on complex tasks. | Streaming Communication in Multi-Agent ReasoningarXiv:2606.05158 · 2026 |
| Coding with "Enemy" | 94% of experienced developers did not notice malicious code — because attention was on the result, not on what exactly the AI did. | Coding with "Enemy": Can Human Developers Detect AI Agent Sabotage?arXiv:2606.05647 · 2026 |
| HUSH-Bench | DeepSeek after such an instruction jumps from 28 to 99.9 out of 100 points for correct behavior. | HUSH-Bench: Measuring Memory-Use Boundaries for Sensitive History in CarXiv:2606.06055 · 2026 |
| CogManip | A system prompt with instructions about objectivity reduces emotional pressure and dependency creation by 1–1.4 points — this has been verified directly on DeepSeek-V3.2. | CogManip: Benchmarking Manipulative Behavior in Multi-Turn InteractionarXiv:2606.06099 · 2026 |
| ToolChoiceConfusion | 83% → 99% accuracy — not from a new prompt and not from changing the model. | ToolChoiceConfusion: Causal Minimal Tool Filtering for Reliable LLM AgarXiv:2606.06284 · 2026 |
| The Cold-Start Safety Gap in LLM Agents | Paradox: an agent is most vulnerable immediately after launch, precisely when the system prompt is fresh and "should work." The warmup method allows raising agent safety by 10–50% without retraining simply through conversation history. | The Cold-Start Safety Gap in LLM AgentsarXiv:2606.07867 · 2026 |
| Silent Failure in LLM Agent Systems | Over 10 steps, an LLM loses 77% of the original data. | Silent Failure in LLM Agent Systems: The Entropy Principle and the InearXiv:2606.08162 · 2026 |
| Sample-Efficient LLM-Based Detection of Malicious Web… | By demonstrating step-by-step solution logic, the method allows reducing the number of necessary examples from 40 to 4 without loss of accuracy, teaching the model methodology rather than surface patterns. | Sample-Efficient LLM-Based Detection of Malicious Web Server Logs witharXiv:2606.08649 · 2026 |
| What Should a Skill Remember? | Compressed the system prompt, spent 14% more tokens. | What Should a Skill Remember? Quality--Cost Trade-offs in Cost-Aware SarXiv:2606.09421 · 2026 |
| Do Vision-Language Models See or Guess? | Researchers removed images altogether, leaving only questions — accuracy dropped to 1–9%. | Do Vision-Language Models See or Guess? Measuring and Reducing TextualarXiv:2606.10400 · 2026 |
| SkillAxe | +28% accuracy after the first iteration — and half the gap to instructions written by human experts is closed. | SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-GuidarXiv:2606.10546 · 2026 |
| Are LLMs Bad at Moral Reasoning? | Instead of "analyze the situation"—you ask to "create evaluation criteria": one semantic shift in the task, and the model covers 83–89% of what experts wrote, plus adds 2.26 times more unique considerations. | Are LLMs Bad at Moral Reasoning?arXiv:2606.11635 · 2026 |
| When the Tool Decides | A weak model (1.5B parameters) agreed with the tool in 60% of cases, a strong one (7B) — in 98%. | When the Tool Decides: LLM Agents Defer Blindly to Graph Neural NetworarXiv:2606.14476 · 2026 |
| Configuration Smells in AGENTS.md Files | Each of the six smells is a documented pattern from 100 real instruction files: bloated context, rare rules in the main prompt, explaining the obvious, blind document references, outdated rules and direct contradictions. | Configuration Smells in AGENTS.md Files: Common Mistakes in ConfigurinarXiv:2606.15828 · 2026 |
| Intelligence Is Not the Bottleneck | The problem is that the same text in the next run gets a different score: variance up to 2.8 points. | Intelligence Is Not the Bottleneck: Validating an LLM First-Pass ManusarXiv:2606.15887 · 2026 |
| The Slop Paradox | Paradox: the request "make it more professional" is more dangerous than "make it shorter"—when cleaning, the text loses fewer words but drifts away from meaning 6–7 times further. | The Slop Paradox: How Synthetic Standardization Erodes Clinical UncertarXiv:2606.17791 · 2026 |
| Evaluating Prompting-Based Defenses Against Domain-… | The success of hidden attacks drops by 55–84%. | Evaluating Prompting-Based Defenses Against Domain-Camouflaged InjectiarXiv:2606.18530 · 2026 |
| Heterogeneous LLM Debate Under Adversarial Peers | 89% of independent revisions in standard multi-agent LLM discussion are harmful: the model doesn't argue, it simply drifts toward someone else's error. | Heterogeneous LLM Debate Under Adversarial Peers: Honest Gains, ReplacarXiv:2606.19826 · 2026 |
| Generative Engine Optimization at Scale | Large brands appear in answers in 73% of cases, niche ones in 11%. | Generative Engine Optimization at Scale: Measuring Brand Visibility AcarXiv:2606.20065 · 2026 |
| Contagion Networks | Different architectures contaminate the result 3–5 times more strongly than one model with different instructions. | Contagion Networks: Evaluator Preference Propagation in Multi-Agent LLarXiv:2606.20493 · 2026 |
| Simulated Customers Never Walk Away | In reality, 25% of disinterested people resist and leave. | Simulated Customers Never Walk Away: Decision Fidelity of LLM User SimarXiv:2606.20708 · 2026 |
| BabelJudge | The bias reaches 25 percentage points, and this is not a glitch in a specific model — this is a documented pattern in all major LLMs. | BabelJudge: Measuring LLM-as-a-Judge Reliability Across Languages and arXiv:2606.22329 · 2026 |
| Text2DSL | The trick is that the dictionary switches the model from guessing to selecting from a list, raising match accuracy from 0.15 to 0.63 and syntactic correctness to 98.6–99.4% without retraining. | Text2DSL: LLM-Based Code Generation for Domain-Specific LanguagesarXiv:2606.22586 · 2026 |
| Capable but Careless | 67.9% is the average percentage of ordinary tasks where AI includes in the answer something that was not asked for: of 15 tested models, 12 leaked unnecessary information in more than half of standard scenarios. | Capable but Careless: Do Computer-Use Agents Follow Contextual IntegriarXiv:2606.23189 · 2026 |
| Why Machines Misread Pedagogical Quality | Result: agreement with humans rises from 55% to 95%. | Why Machines Misread Pedagogical Quality: Human-Machine Alignment in LarXiv:2606.23629 · 2026 |
| Accuracy and Satisfaction in Multi-Turn LLM Dialogues for… | Professional developers agreed with GitHub Copilot answers in 91–94% of cases. | Accuracy and Satisfaction in Multi-Turn LLM Dialogues for NFR AssessmearXiv:2606.24834 · 2026 |
| How Large Language Models Source Brand Reputation Across… | 85.7% of what AI says about your brand was not written by you. | How Large Language Models Source Brand Reputation Across Languages andarXiv:2606.25787 · 2026 |
| Can LLMs Judge Better Than They Generate? | Found: when checking its own answer, LLM looks at it for 0.3% of the time. | Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, MarXiv:2606.28050 · 2026 |
| SafeGEO | This instruction alone cuts the manipulation effect by 39%. | SafeGEO: Understanding Generative Engine Optimization Risks in RecommearXiv:2606.28356 · 2026 |
| LEDGER | A model at maximum reasoning power without a map loses to the same model on simple settings — but with an explicit dependency map: 56% versus 76% consistency. | LEDGER: Scaling Agentic Document Editing with Dependency-aware Graph RarXiv:2606.28379 · 2026 |
| Hierarchical Experimentalist Agents | The bank grows, the best records go to the beginning of the next chat — +75% success on the most difficult tasks. | Hierarchical Experimentalist AgentsarXiv:2606.29315 · 2026 |
| SurrogateShield | The model sees a complete text and produces a complete answer — 94.85% semantic accuracy versus 81.59% with placeholders. | SurrogateShield: Beyond Redaction for High-Utility, Privacy-PreservingarXiv:2606.29567 · 2026 |
| How LLMs See Creativity | The trick: two mandatory elements of the prompt — "use the ENTIRE scale, including extreme values" and "evaluate the idea, not the execution" — without them the model clings to 3-4 and penalizes drafts automatically. | How LLMs See Creativity: Zero-Shot Scoring of Visual Creativity with IarXiv:2606.29672 · 2026 |
| Entity Binding Failures in Tool-Augmented Agents | 24–26% of AI agent actions with tools hit the right tool but the wrong target — and the model stays silent about it. | Entity Binding Failures in Tool-Augmented AgentsarXiv:2606.30531 · 2026 |
| Guiding Human Validation of LLM-Generated Code via… | 43–47% of errors in LLM-generated code are your errors, not the model's. | Guiding Human Validation of LLM-Generated Code via Verifiable LiteratearXiv:2607.02333 · 2026 |
| Beyond the Leaderboard | GPT-4 handles each planning step separately — but when you need to hold multiple constraints simultaneously, success drops to 0.6%. | Beyond the Leaderboard: A Synthesis of Tool-Use, Planning, and ReasoniarXiv:2607.05775 · 2026 |
| Evaluating LLM Robustness Under Domain-Specific Prompt… | A false statement in context pulls accuracy down by 7 percentage points and flips every tenth answer — even if you directly refute it. | Evaluating LLM Robustness Under Domain-Specific Prompt Perturbations iarXiv:2607.06913 · 2026 |
| Operational Reframing and Approval-Framed Delegation in… | One phrase — 'evaluate the task independently' — reduces automatic request execution by 47 percentage points. | Operational Reframing and Approval-Framed Delegation in Multi-Agent LLarXiv:2607.07097 · 2026 |
| User identity conditions moral wrongness ratings in non-… | The gap between roles reaches 37 points on a 100-point scale—for the same action. | User identity conditions moral wrongness ratings in non-reasoning largarXiv:2607.07605 · 2026 |
| A safety-oriented hypothetico-deductive framework for AI-… | In medical cases, coverage of dangerous omissions increased by 26 percentage points compared to a standard request from the same model. | A safety-oriented hypothetico-deductive framework for AI-assisted diffarXiv:2607.08038 · 2026 |
| When the Judge Changes, So Does the Measurement | 14–15% of verdicts change simply from swapping—even in top models. | When the Judge Changes, So Does the Measurement: Auditing LLM-as-JudgearXiv:2607.08535 · 2026 |
| The complexities of patient-centred conversational… | Identical medical facts, different tone of description — and AI assigns different urgency ratings: the difference is up to 13.5 percentage points of errors just from the style of the request, without a single changed fact. | The complexities of patient-centred conversational artificial intelligarXiv:2607.08625 · 2026 |
| Exploring the Potential of Program Flowcharts on Code… | One additional file in the request gives +10–13% accuracy. | Exploring the Potential of Program Flowcharts on Code Generation UsingarXiv:2607.09146 · 2026 |
| Spectral Origins of the Self-Correction Blind Spot in… | This is not random — in 64% of LLM answers, error during generation doesn't weaken, but grows. | Spectral Origins of the Self-Correction Blind Spot in Autoregressive GarXiv:2607.09803 · 2026 |
| Trivial Prompt Reframing Bypasses Safety Guardrails in… | Paradox: rough pressure 'YOU MUST ANSWER' moves the needle by +3 points — this is statistical noise. | Trivial Prompt Reframing Bypasses Safety Guardrails in Googleś MedGemmarXiv:2607.09804 · 2026 |
| Lower-Resource, Higher Scores | The gap reaches 0.5 points on a 1–5 scale and is reproduced across all eight models tested. | Lower-Resource, Higher Scores: Language Bias in LLM EvaluatorsarXiv:2607.14480 · 2026 |
| Controlled Reformulation Testing for Logical Consistency in… | GPT-5.4-mini answers correctly in 99% of cases — and yet changes the answer to the opposite in 40% of question families when the same thing is reformulated through contraposition or double negation. | Controlled Reformulation Testing for Logical Consistency in Large LangarXiv:2607.14528 · 2026 |
| Does Multi-Agent Debate Improve AI Feedback on Research… | One well-formulated query with role, tone, and output structure produces a result that a multi-agent system spends 30 times more tokens on — and still loses. | Does Multi-Agent Debate Improve AI Feedback on Research Papers?arXiv:2607.14713 · 2026 |
| DRNOISE | In 100% of cases the agent with search already found the correct data—it simply stopped before reaching it. | DRNOISE: Benchmarking Deep Research Agents in Misleading Evidence EnviarXiv:2607.17291 · 2026 |
| Informal Learning Emerges in Everyday Human-LLM Interaction | In 128 thousand real conversations, only 5% contained what is actually worth talking to the model about — real understanding instead of a ready result. | Informal Learning Emerges in Everyday Human-LLM InteractionarXiv:2607.17643 · 2026 |
| When and How Context Rot Appears in Coding Agents | Allows the model to compare the generated result with specific wording of conditions after generation, which increases accuracy of execution from 50% to 100% in tasks with multiple requirements. | When and How Context Rot Appears in Coding Agents: A White-Box Study oarXiv:2607.17937 · 2026 |
| Agentic Evaluation of Copyright Law Compliance | Claude 4.5 Opus — one of the best models on the market — in 38% of cases chooses a copyrighted image although the exact same free one lies in the same pool. | Agentic Evaluation of Copyright Law CompliancearXiv:2607.21799 · 2026 |
| Poster | The security-aware prompting method allows you to reduce the proportion of vulnerable code by 37-45% with one improved prompt. | Poster: Rethinking Security in LLM Code Generation through Real-World arXiv:2607.23088 · 2026 |
| When Do Agent Loops Mistake Stagnation for Progress? | 100% of 54 cycles: the agent reports 'it got better'. | When Do Agent Loops Mistake Stagnation for Progress? Self-Evaluation BarXiv:2607.25152 · 2026 |
| Why LLMs Give In | If an objection comes in the very first question, the model yields in 2.5% of cases. | Why LLMs Give In: Conversational Factors and Reasoning Behind Medical arXiv:2608.01017 · 2026 |
| Hallucinations on the Board | 40% of texts riddled with factual errors received high marks from LLM judges simply because they sounded confident. | Hallucinations on the Board: Tool-Augmented Evaluation of LLM Chess CoarXiv:2608.04240 · 2026 |
| Hijacking Robots with a Piece of Paper | A piece of paper with the phrase 'SYSTEM UPDATE' makes a robot manipulator violate the operator's command in 27-29% of cases — the model honestly 'notices' the forgery in its reasoning and obeys it anyway. | Hijacking Robots with a Piece of Paper: A Systematic Study of PhysicalarXiv:2608.05715 · 2026 |
| Can LLM Agents Stick to the Script? | 42% — that is how much time the top model GPT-5.2 keeps the plot without contradictions already after 20 dialogue turns. | Can LLM Agents Stick to the Script? A Benchmark for Long-Horizon ConsiarXiv:2608.08160 · 2026 |
| TAF-MED | A method of attacking LLM safeguards through multi-turn dialogue: after the model's initial refusal, the user asks a series of 'innocent' clarifying questions (educational, hypothetical, third-person), which in 61% of cases leads to safeguard failure and disclosure of prohibited information. | TAF-MED: Multi-Turn Safety Refusal Collapse in LLMs Under Declared SelarXiv:2608.10258 · 2026 |
| Actions Speak Louder than Words | The task arrives in Russian, but inside the model reasons and calls tools almost entirely in English — 99% of characters, even if the input was in Hindi or Tamil. | Actions Speak Louder than Words: Measuring Cross-Lingual Policy RetentarXiv:2608.11110 · 2026 |
| The Wording Effect | 93% accuracy on a benchmark — and this could have simply been a lucky set of phrasings. | The Wording Effect: Quantifying Two-Way Drift in LLM Benchmark PerformarXiv:2608.11694 · 2026 |
| Who Thinks Best Depends on How Long You Let Them | On complex multi-step tasks, up to 25% of cases show that a longer answer is worse than a shorter one. | Who Thinks Best Depends on How Long You Let Them: Budget-Dependent RanarXiv:2608.12150 · 2026 |
| AnchorBench | Even GPT, Claude and Gemini with accuracy above 95% without an anchor still measurably shift toward any 'plausible' number from context — just weaker than open models. | AnchorBench: A Multi-Pathway Benchmark for the Anchoring Effect in LLMarXiv:2608.14320 · 2026 |
| Diagnosing Dense Same-Class Attribute Misbinding in Large… | In 80% of confusion cases the model pulls the color from a neighboring object standing closely nearby — not from a random object in the photo. | Diagnosing Dense Same-Class Attribute Misbinding in Large Vision-LanguarXiv:2608.16805 · 2026 |
| Quantifying non deterministic drift in large language models | In practice, GPT-4o-mini produces a new variant in almost every fourth run (~24%), even with an identical prompt. | Quantifying non deterministic drift in large language modelsarXiv:2601.19934 · 2026 |
| Chain-in-Tree | Result: 10-20 times slower than simple approaches, burning tokens on trivial sections. | Chain-in-Tree: Back to Sequential Reasoning in LLM Tree SearcharXiv:2509.25835 · 2025 |
| Multi-Modal Fact-Verification Framework for Reducing… | The method reduces hallucinations by 67% through an external verification system. | Multi-Modal Fact-Verification Framework for Reducing Hallucinations inarXiv:2510.22751 · 2025 |
| Retrieval-Augmented Generation (RAG) | LLM without textbook context fails on school questions: F1 drops from 0.89 (general knowledge) to 0.39 (NCERT curriculum). | PustakAI: Curriculum-Aligned and Interactive Textbooks Using Large LanarXiv:2511.10002 · 2025 |
| Retrieval-Augmented Generation (RAG) | GPT-4o with web search finds relevant fact-checking sources in only 50% of cases, and reasoning models like o1 provide only +6% accuracy improvement — the problem is not HOW models think, but WHETHER they have the correct information. | Large Language Models Require Curated Context for Reliable Political FarXiv:2511.18749 · 2025 |
| Too Late to Recall | 11 of 14 tested VLMs fail on facts when working with images — accuracy drops by 6–44%. | Too Late to Recall: Explaining the Two-Hop Problem in Multimodal KnowlarXiv:2512.03276 · 2025 |
| Catching UX Flaws in Code | It was found that LLMs consistently identify WHAT is wrong — 84% agreement between runs. | Catching UX Flaws in Code: Leveraging LLMs to Identify Usability FlawsarXiv:2512.04262 · 2025 |
| Evaluating Long-Context Reasoning in LLM-Based WebAgents | Long dialogue kills LLM memory: at 25k-150k tokens, approximately 50-300 messages, task success drops from 40-50% to less than 10%. | Evaluating Long-Context Reasoning in LLM-Based WebAgentsarXiv:2512.04307 · 2025 |
| How Do LLMs Fail In Agentic Scenarios? | Paradox: a model with 400 billion parameters performs only 16% better than a model with 32 billion on multi-step tasks (reading files, SQL queries, CSV analysis). | How Do LLMs Fail In Agentic Scenarios? A Qualitative Analysis of SuccearXiv:2512.07497 · 2025 |
| Secure or Suspect? | Research showed: lightweight models (quantized — compressed to save memory) hallucinate up to 98% of packages versus 40-50% in originals. | Secure or Suspect? Investigating Package Hallucinations of Shell CommaarXiv:2512.08213 · 2025 |
| DeepSeek's WEIRD Behavior | Cultural prompts in English are most effective for non-western cultures, providing up to +29.8% improvement in cultural alignment. | DeepSeek's WEIRD Behavior: The cultural alignment of Large Language MoarXiv:2512.09772 · 2025 |
| Persistent Personas? | You give the model a character (vampire Lestat, editor Ilyahov, creative marketer) — after 50-100 messages you get an ordinary model without character. | Persistent Personas? Role-Playing, Instruction Following, and Safety iarXiv:2512.12775 · 2025 |
| Us-vs-Them bias in Large Language Models | Exploratory thinking forces the model to consider multiple perspectives before conclusion, reducing bias by 60% for own and 54% for others. | Us-vs-Them bias in Large Language ModelsarXiv:2512.13699 · 2025 |
| Are We on the Right Way to Assessing LLM-as-a-Judge? | When you ask an LLM to choose the best of three text variants, it changes its opinion in 76% of cases simply due to reordering, so positional bias is real. | Are We on the Right Way to Assessing LLM-as-a-Judge?arXiv:2512.16041 · 2025 |
| Needle in the Web | Accuracy of such queries is below 35%. | Needle in the Web: A Benchmark for Retrieving Targeted Web Pages in tharXiv:2512.16553 · 2025 |
| Visually Prompted Benchmarks Are Surprisingly Fragile | You change red to blue or reduce the marker size – accuracy drops 10-21%. | Visually Prompted Benchmarks Are Surprisingly FragilearXiv:2512.17875 · 2025 |
| DREAM | Each step seems simple, but together they produce a complex result—70%+ success in chains versus failure of single requests. | DREAM: Dynamic Red-teaming across Environments for AI ModelsarXiv:2512.19016 · 2025 |
| Grading Handwritten Engineering Exams with Multimodal Large… | The method allows obtaining calibrated objective evaluations of texts, contractor work, ideas — with accuracy of approximately 8 points from expert assessment. | Grading Handwritten Engineering Exams with Multimodal Large Language MarXiv:2601.00730 · 2026 |
| GAMBIT | Paradox: reasoning models (GPT-4o with chain of thought) turned out to be MORE vulnerable to filter bypass than simple models—92% successful attacks versus 31%. | GAMBIT: A Gamified Jailbreak Framework for Multimodal Large Language MarXiv:2601.03416 · 2026 |
| Large Language Models Are Bad Dice Players | When increasing the sample from 1000 to 2000, the distance from the target distribution grows instead of falling. | Large Language Models Are Bad Dice Players: LLMs Struggle to Generate arXiv:2601.05414 · 2026 |
| Evaluating Robustness of Large Language Models in… | ServiceNow tested 11 models (from 4B to 120B parameters) for resistance to five types of changes: spacing, punctuation, section rearrangement, output format change (JSON/XML/YAML), multilingual inputs. | Evaluating Robustness of Large Language Models in Enterprise ApplicatiarXiv:2601.06341 · 2026 |
| BenchOverflow | Allows reducing generated text volume 6-10 times (from 3000 to 300-500 tokens) without loss of informativeness, switching the model from comprehensive topic coverage mode to concise summary mode. | BenchOverflow: Measuring Overflow in Large Language Models via Plain-TarXiv:2601.08490 · 2026 |
| Contextual StereoSet | The same model produces 12 percentage points more stereotypes if you mention 1990 instead of 2030 in the prompt, or 'someone mentioned' instead of 'I officially declare'. | Contextual StereoSet: Stress-Testing Bias Alignment Robustness in LargarXiv:2601.10460 · 2026 |
| A Concise Agent is Less Expert | Expertise drops from 50% to 25.6%, usefulness to 23.2%. | A Concise Agent is Less Expert: Revealing Side Effects of Using Style arXiv:2601.10809 · 2026 |
| Intelligence Degradation in Long-Context LLMs | Qwen2.5-7B loses 45.5% performance when moving from 51K to 64K tokens — F1 score drops from 0.55 to 0.30. | Intelligence Degradation in Long-Context LLMs: Critical Threshold DetearXiv:2601.15300 · 2026 |
| CogToM | Paradox: LLM analyzes hidden emotions like an experienced psychologist (96% accuracy on a task for 6-year-olds) but fails at "what does the person across from you see" (20% on a task for 3-year-olds). | CogToM: A Comprehensive Theory of Mind Benchmark inspired by Human CogarXiv:2601.15628 · 2026 |
| Regional Bias in Large Language Models | You ask 'which of two candidates is better?' (both are equal in experience) — GPT-3.5 in 95% of cases will name a country and invent a reason. | Regional Bias in Large Language ModelsarXiv:2601.16349 · 2026 |
| MortalMATH | The model solves an algebraic equation with 95% accuracy while ignoring the user's phrase 'My parachute didn't open, I'm falling' — and spends 10-15 seconds on this. | MortalMATH: Evaluating the Conflict Between Reasoning Objectives and EarXiv:2601.18790 · 2026 |
| Counterfactual Cultural Cues Reduce Medical QA Accuracy in… | GPT-5.2 drops from 92.67% to 85% when 'Muslim from a remote community' appears in the prompt—minus 7 percentage points from a clinically irrelevant detail. | Counterfactual Cultural Cues Reduce Medical QA Accuracy in LLMs: IdentarXiv:2601.20102 · 2026 |
| SHARP | Paradox: models with identical average quality (risk 0.45 vs 0.46) showed in the worst 5% of answers average harm of 3.50 versus 3.78—these are qualitatively different risk profiles. | SHARP: Social Harm Analysis via Risk Profiles for Measuring InequitiesarXiv:2601.21235 · 2026 |
| Feedback by Design | But with LLM this skill breaks—out of a million ChatGPT dialogs only 3.89% contain feedback, and most of it is useless 'wrong' or 'try again'. | Feedback by Design: Understanding and Overcoming User Feedback BarrierarXiv:2602.01405 · 2026 |
| CIPHER | Paradox: you ask an LLM to 'make secure code' — you get 71.7% vulnerable code. | CIPHER: Cryptographic Insecurity Profiling via Hybrid Evaluation of RearXiv:2602.01438 · 2026 |
| Seeing Is Believing? | People handle visual illusions at 93.3%, the best model (OpenAI o3)—only at 69.2%. | Seeing Is Believing? A Benchmark for Multimodal Large Language Models arXiv:2602.01816 · 2026 |
| A Positive Case for Faithfulness | Self-explanations improve prediction of model behavior by 11-37% compared to external interpretations. | A Positive Case for Faithfulness: LLM Self-Explanations Help Predict MarXiv:2602.02639 · 2026 |
| Precision in Practice | Paradox: GPT-4.5 describes functions semantically correctly, but industrial developers reject 57.4% of descriptions. | Precision in Practice: Knowledge Guided Code Summarizing Grounded in IarXiv:2602.03400 · 2026 |
| Can Developers rely on LLMs for Secure IaC Development? | Especially on large scripts: GPT-4o and Gemini find only 42-51% of issues (missing hard passwords in code, weak encryption, open ports). | Can Developers rely on LLMs for Secure IaC Development?arXiv:2602.03648 · 2026 |
| Do Vision-Language Models Respect Contextual Integrity in… | GPT-5 provides precise location in 47.6% of cases where a person would say 'I cannot answer'. | Do Vision-Language Models Respect Contextual Integrity in Location DisarXiv:2602.05023 · 2026 |
| Stop Testing Attacks, Start Diagnosing Defenses | Paradox: LLM refuses formally, but leaks 30-70% of requested information. | Stop Testing Attacks, Start Diagnosing Defenses: The Four-Checkpoint FarXiv:2602.09629 · 2026 |
| When Audio-LLMs Don't Listen | You speak one thing with your voice in ChatGPT Voice, but something different is written in the chat history — the model ignores your voice in 16.6% of cases (Gemini 2.0 Flash). | When Audio-LLMs Don't Listen: A Cross-Linguistic Study of Modality ArbarXiv:2602.11488 · 2026 |
| BaziQA-Benchmark | Accuracy in time-based forecasts +8-10%, but on simple tasks it can hurt up to -20%. | BaziQA-Benchmark: Evaluating Symbolic and Temporally Compositional ReaarXiv:2602.12889 · 2026 |
| Evaluating Prompt Engineering Techniques for RAG in Small… | Small models struggle on multi-hop questions — when you need to connect facts from 3-4 sources, they get lost in context and cannot identify the main point. | Evaluating Prompt Engineering Techniques for RAG in Small Language ModarXiv:2602.13890 · 2026 |
| The Potential of CoT for Reasoning | A paradox: reasoning models (o1, R1, Qwen3) get stuck 2 times more often than regular models. | The Potential of CoT for Reasoning: A Closer Look at Trace DynamicsarXiv:2602.14903 · 2026 |
| Counterfactual Fairness Evaluation of LLM-Based Contact… | Research revealed systematic bias in 18 models across 13 dimensions: names influence decisions in 13-16% of cases. | Counterfactual Fairness Evaluation of LLM-Based Contact Center Agent QarXiv:2602.14970 · 2026 |
| AIDG | 7.75 times — that is the gap between a blind question and a specific hypothesis. | AIDG: A Formal Decomposition of Information Extraction and ContainmentarXiv:2602.17443 · 2026 |
| Story Point Estimation Using Large Language Models | A zero-shot approach to effort estimation through an LLM that outperforms ML-models trained on 80% of historical data. | Story Point Estimation Using Large Language ModelsarXiv:2603.06276 · 2026 |
| The Anatomy of Uncertainty in LLMs | With complete apparent confidence, an LLM makes mistakes in 17% of cases. | The Anatomy of Uncertainty in LLMsarXiv:2603.24967 · 2026 |
| RubricEval | Accuracy increases by 7–12 percentage points — even GPT-4o on complex cases previously made mistakes in almost every other case. | RubricEval: A Rubric-Level Meta-Evaluation Benchmark for LLM Judges inarXiv:2603.25133 · 2026 |
| SenseMath | Add one instruction—'look for smart computational paths'—and the model switches from 'executing calculator' mode to 'smart solver' mode: the frequency of smart strategies grows from 40% to 68–86%, accuracy on complex tasks—up to +15 percentage points. | SenseMath: Do LLMs Have Number Sense? Evaluating Shortcut Use, JudgmenarXiv:2604.01988 · 2026 |
| FreakOut-LLM | The scenario "crisis/panic/deadline" in the system prompt shifts probabilities: growth of violations from 1.6% to 2.6%—this is +65% relative to the baseline. | FreakOut-LLM: The Effect of Emotional Stimuli on Safety AlignmentarXiv:2604.04992 · 2026 |
| Self-Preference Bias in Rubric-Based Evaluation of Large… | When you ask a model to evaluate text that it wrote itself—it 'does not notice' requirement violations 50% more often than when evaluating someone else's. | Self-Preference Bias in Rubric-Based Evaluation of Large Language ModearXiv:2604.06996 · 2026 |
| Do LLMs Build Spatial World Models? | The same maze, the same data — but in ASCII-diagram format the model solves it 2–5 times worse than in the format of a list «A connects to B». | Do LLMs Build Spatial World Models? Evidence from Grid-World Maze TaskarXiv:2604.10690 · 2026 |
| Calibrated Confidence Estimation for Tabular Question… | "I am 99% confident" — is correct in 65–76% of cases. | Calibrated Confidence Estimation for Tabular Question AnsweringarXiv:2604.12491 · 2026 |
| Evaluating Answer Leakage Robustness of LLM Tutors against… | It was specifically context reformulation (contextual manipulation) that turned out to be the most effective pressure technique on LLM with any role — the average percentage of instruction violations across all tested models was 74%. | Evaluating Answer Leakage Robustness of LLM Tutors against AdversarialarXiv:2604.18660 · 2026 |
| Beyond Semantic Similarity | 51% similarity with the reference answer — and at the same time 6% accuracy for specific medications. | Beyond Semantic Similarity: A Component-Wise Evaluation Framework for arXiv:2604.19281 · 2026 |
| Shift-Up | 52% of prompts with the usual approach is not development but cleanup. | Shift-Up: A Framework for Software Engineering Guardrails in AI-nativearXiv:2604.20436 · 2026 |
| From If-Statements to ML Pipelines | The model neatly dropped 'favorite color' from 89% of analyses, understanding that the attribute is meaningless. | From If-Statements to ML Pipelines: Revisiting Bias in Code-GenerationarXiv:2604.21716 · 2026 |
| From 0-Order Selection to 2-Order Judgment | GLM-4.7 scored 83% accuracy on regular questions — and showed the worst gap among twelve models on compound conditions. | From 0-Order Selection to 2-Order Judgment: Combinatorial Hardening ExarXiv:2605.07268 · 2026 |
| Unsolvability Ceiling in Multi-LLM Routing | The judge-LLM gives 8/10 to a beautiful incorrect answer and 4/10 to a correct concise one — the discrepancy with actual accuracy reaches 24 percentage points. | Unsolvability Ceiling in Multi-LLM Routing: An Empirical Study of EvalarXiv:2605.07395 · 2026 |
| Why Retrying Fails | The number after which you stop hitting 'redo' in an old chat: the probability of an error on the second attempt is 7 times higher than the first — if the first failure remained in the context. | Why Retrying Fails: Context Contamination in LLM Agent PipelinesarXiv:2605.08563 · 2026 |
| Beyond Accuracy | 98% accuracy with one solution — and only 50% coverage of possible strategies. | Beyond Accuracy: Evaluating Strategy Diversity in LLM Mathematical ReaarXiv:2605.09292 · 2026 |
| Visual Aesthetic Benchmark | When you ask AI to 'rate from 1 to 10' you get a number. | Visual Aesthetic Benchmark: Can Frontier Models Judge Beauty?arXiv:2605.12684 · 2026 |
| LLM-Based Persuasion Enables Guardrail Override in Frontier… | Research shows: the same Claude Opus 4.7 that refuses a direct request fulfills it in 65% of cases if you talk to it in a structured way over five turns. | LLM-Based Persuasion Enables Guardrail Override in Frontier LLMsarXiv:2605.13334 · 2026 |
| Edit-level Majority Voting Mitigates Over-Correction in… | The trick: the model's inconsistency itself becomes a filter — run the text 5 times, extract edits from each version, keep only those that appeared in most attempts. | Edit-level Majority Voting Mitigates Over-Correction in LLM-based GramarXiv:2605.13624 · 2026 |
| Evaluating the Utility of Personal Health Records in… | Temporal errors decrease by approximately 98% without code and technical settings. | Evaluating the Utility of Personal Health Records in Personalized HealarXiv:2605.18937 · 2026 |
| Evaluating Large Language Models as Live Strategic Agents | In a 32-game tournament between Gemini, GPT, Claude and Kimi in the strategic game Risk this gave +27% of conquered territories—without changing the model, without additional training. | Evaluating Large Language Models as Live Strategic Agents: Provider PearXiv:2605.22238 · 2026 |
| SemanticZip | Result: 40–46% fewer tokens, zero repeated explanations. | SemanticZip: A Pilot Framework for Lossy Text Compression with LLMs asarXiv:2605.24541 · 2026 |
| SLMJury | For judgments the criterion is fuzzy, step-by-step reasoning structures the search by considerations, and here the gain reaches +23% accuracy. | SLMJury: Can Small Language Models Judge as Well as Large Ones?arXiv:2606.07810 · 2026 |
| Measuring Whether LLM Tutors Teach or Solve | Ability to solve does not mean ability to teach: among 8 AI-tutors the correlation between these abilities is only 0.42, and the gap in pedagogical metrics between the best and the worst is 45 points. | Measuring Whether LLM Tutors Teach or Solve: A Diagnostic for EducatioarXiv:2606.16206 · 2026 |
| ReContext | The trick: we do not remove the unnecessary, but duplicate the important — the full document remains, but 128 'golden tokens' (in which 50–80% of the answer resides) stand directly before the question at the moment of generation. | ReContext: Recursive Evidence Replay as LLM Harness for Long-Context RarXiv:2607.02509 · 2026 |
| Prompting Beats Fine-Tuning | Add a paragraph around the evaluated fragment — and ranking accuracy drops in half: NDCG from 0.53 to 0.28. | Prompting Beats Fine-Tuning: Generative Expected Value Scoring for StaarXiv:2607.05582 · 2026 |
| Programmers Are Poor and Overconfident Judges of LLM-… | You correctly accept a correct AI answer in 74% of cases. | Programmers Are Poor and Overconfident Judges of LLM-Generated AssertiarXiv:2607.08885 · 2026 |
| The Paternalistic Filter | Refusals — 14.5 percentage points higher. | The Paternalistic Filter: Epistemic Injustice and Differential RefusalarXiv:2607.11292 · 2026 |
| Relational Positioning as a Measurable Risk Object | The research gives two specific levers: set the tone in the first 6 turns while the position is not frozen, and erase fabricated biographies with one phrase (self-confabulation drops from 40% to ~1% with one line in the prompt). | Relational Positioning as a Measurable Risk Object: History-Carried LoarXiv:2607.11437 · 2026 |
| Not Birds of a Feather | 4 times out of 375 trials — that is how many times the model chose a friendly or sociable partner, although these are exactly the traits that drive real teams of people. | Not Birds of a Feather: Personality-Based Partner Selection in LLM AgearXiv:2607.19785 · 2026 |
| Frontier Financial Judgement | Even the best model guesses all three criteria correctly in only 50% of cases. | Frontier Financial Judgement: Can agents tell what might move a stock?arXiv:2607.20645 · 2026 |
| The Prompt Is Not the Query | The last message in a dialogue with an LLM contains only 35% of the unique words from the entire conversation. | The Prompt Is Not the Query: How Request State Evolves Across Multi-TuarXiv:2607.22392 · 2026 |
| A Taxonomy of Confabulations and the Perception-Reality Gap… | Cambridge researchers tested voice editing of VR scenes through LLM: in complex scenes people missed up to 40% of actual model errors. | A Taxonomy of Confabulations and the Perception-Reality Gap in LLM-AssarXiv:2607.23213 · 2026 |
| Ranked by Position | 57% of initially irrelevant candidates make it to the top-5—just rearrange the resumes, without changing a word in the content. | Ranked by Position: Order Sensitivity as an Exploitable Attack SurfacearXiv:2607.24869 · 2026 |
| Two Calls Beat Five Agents | It was only necessary to ask 'double-check and give the final version' — accuracy dropped to 66%. | Two Calls Beat Five Agents: Evaluating Multi-Agent Pipelines Against SarXiv:2607.26922 · 2026 |
| Single Canonical Prompts Underestimate LLM Safety's… | The share of risky responses grows 1.3–2.2 times compared to checking a single formulation. | Single Canonical Prompts Underestimate LLM Safety's Surface-Form SensiarXiv:2608.02665 · 2026 |
| Distractor-Aware Truncation | Deleted 75% of text—and in 99% of cases the answer itself went with it. | Distractor-Aware Truncation: Disentangling Context-Length Effects fromarXiv:2608.03297 · 2026 |
| SearchAuditor | 27% of all search agent errors are simple confusion: the agent found the correct answer then lost it among similar candidates. | SearchAuditor: Auditing and Attributing Failures in Long-Horizon SearcarXiv:2608.05212 · 2026 |
| Persona Conditioning as an Assessor-Sensitivity Probe for… | The method makes it possible to find hidden ambiguity in content: ask the model to evaluate one text from 3-5 different roles and compare the scores. | Persona Conditioning as an Assessor-Sensitivity Probe for LLM-Based IRarXiv:2608.10385 · 2026 |
| TRACES | Discovered: LLMs agree to develop pseudoscientific articles in 93% of cases — they recognize only a couple of 'famous' scandals by keywords. | TRACES: A Benchmark for Epistemic Reliability in Scientific Reasoning arXiv:2608.11415 · 2026 |
| Self-Training for Conciseness (FS-BoN) | Simple prompt commands like 'Be concise' work unreliably and often reduce accuracy, while targeted fine-tuning on your own concise answers allows you to consistently reduce answer length by ~30% without loss of quality. | Self-Training Elicits Concise Reasoning in Large Language ModelsarXiv:2502.20122 · 2025 |
| An Empirical Study of Multi-Agent RAG for Real-World… | Discovered: LLMs are catastrophically unreliable in tasks requiring precise facts — standard model produced 15% hallucinations when consulting applicants. | An Empirical Study of Multi-Agent RAG for Real-World University AdmissarXiv:2507.11272 · 2025 |
| RepoTransAgent | The result improved from 26-30% success to 46-55%, almost twice as good. | RepoTransAgent: Multi-Agent LLM Framework for Repository-Aware Code TrarXiv:2508.17720 · 2025 |
| Joint Enhancement of Relational Reasoning for Long-Context… | Reduces hallucinations and increases accuracy from 34% to 67% on long documents. | Joint Enhancement of Relational Reasoning for Long-Context LLMsarXiv:2508.20351 · 2025 |
| Mind the Gap | Tokenization optimization technique for multiple-choice tasks: combining a space with an answer letter into a single token (␣A instead of ␣+A) increases accuracy by 11% and eliminates instability in LLM predictions. | Mind the Gap: A Closer Look at Tokenization for Multiple-Choice QuestiarXiv:2509.15020 · 2025 |
| LLM Microscope | A classifier based on internal activations predicts correctness with 75% accuracy, but the model itself systematically overestimates confidence. | LLM Microscope: What Model Internals Reveal About Answer Correctness aarXiv:2510.04013 · 2025 |
| Harmonizing Diverse Models | Increases response consistency by 47.5%. | Harmonizing Diverse Models: A Layer-wise Merging Strategy for ConsistearXiv:2510.14915 · 2025 |
| Teaming LLMs to Detect and Mitigate Hallucinations | Increases accuracy by 15-30% compared to Self-Consistency. | Teaming LLMs to Detect and Mitigate HallucinationsarXiv:2510.19507 · 2025 |
| Remote Labor Index | Paradox: AI almost at human level on tests—and 2.5% of closed projects in reality. | Remote Labor Index: Measuring AI Automation of Remote WorkarXiv:2510.26787 · 2025 |
| Thinking Forward and Backward | Two rewards plus a cascading structure (early steps are more important) give +18.2% accuracy while using a quarter of the data for training. | Thinking Forward and Backward: Multi-Objective Reinforcement Learning arXiv:2511.09109 · 2025 |
| Normalized Contextual Calibration (NCC) | Increases F1-score by 10% for datasets with long classes. | Mitigating Label Length Bias in Large Language ModelsarXiv:2511.14385 · 2025 |
| PARROT | GPT-4 (old version) was right in 72% of cases, after the phrase 'As an experienced doctor, I am confident that...'—only 18%. | PARROT: Persuasion and Agreement Robustness Rating of Output Truth -- arXiv:2511.17220 · 2025 |
| Automated Risk-of-Bias Assessment of Randomized Controlled… | LLM loses up to 20 percentage points of accuracy when trying to search for facts and evaluate them in one question. | Automated Risk-of-Bias Assessment of Randomized Controlled Trials: A FarXiv:2512.01452 · 2025 |
| AsymPuzl | GPT-4o with feedback 'both parts solved + details of both' shows 57% success. | AsymPuzl: An Asymmetric Puzzle for multi-agent cooperationarXiv:2512.03466 · 2025 |
| Replicating TEMPEST at Scale | Scale paradox: Mistral Large with 675 billion parameters is cracked 100% of the time on the first attempt, while MiniMax at 230 billion — only in 55% of cases over 23 turns. | Replicating TEMPEST at Scale: Multi-Turn Adversarial Attacks Against TarXiv:2512.07059 · 2025 |
| Can GPT replace human raters? | But the model fails on metaphors about the body and movement (correlation drops to 0.37) — it has no physical experience, works only with text. | Can GPT replace human raters? Validity and reliability of machine-genearXiv:2512.12444 · 2025 |
| Large Language Newsvendor | Paradox: GPT-4 deviates from optimum 70% more than humans in inventory planning tasks. | Large Language Newsvendor: Decision Biases and Cognitive MechanismsarXiv:2512.12552 · 2025 |
| CTIGuardian | Problem: Models fine-tuned on private data memorize and output sensitive information — through selected prefixes, researchers extracted 12-21% of emails, IP addresses, phone numbers. | CTIGuardian: A Few-Shot Framework for Mitigating Privacy Leakage in FiarXiv:2512.12914 · 2025 |
| ChatGPT and Gemini participated in the Korean College… | It shows that input structure critically affects accuracy: separating text and visual content increases results from 8-28% to 52-80%. | ChatGPT and Gemini participated in the Korean College Scholastic AbiliarXiv:2512.15298 · 2025 |
| Step-GUI Technical Report | Evaluation by final result instead of step-by-step checking ensures objectivity >90% and saves resources 10-100 times over. | Step-GUI Technical ReportarXiv:2512.15431 · 2025 |
| Task Schema and Binding | Research through neuron activation substitution revealed physical separation: Task Schema (task type 'name → profession') lives in late MLP-layers and transfers at 100%, while Binding (specific pairs 'Anna → doctor') lives in accumulated activations and transfers in only 62% of cases. | Task Schema and Binding: A Double Dissociation Study of In-Context LeaarXiv:2512.17325 · 2025 |
| Steerability of Instrumental-Convergence Tendencies in LLMs | Add 'no workarounds', 'transparency is mandatory', 'compliance is more important than speed' — behavior changes from 82% attempted tricks to 3%. | Steerability of Instrumental-Convergence Tendencies in LLMsarXiv:2601.01584 · 2026 |
| Project Ariadne | In 77% of cases the model arrives at an identical conclusion when a directly opposite statement is inserted into the reasoning chain. | Project Ariadne: A Structural Causal Framework for Auditing FaithfulnearXiv:2601.02314 · 2026 |
| Ethical Risks in Deploying Large Language Models | In narrow domains (medical ethics, professional contexts) protection collapsed in 5 out of 7 models, 96-100% breach rate. | Ethical Risks in Deploying Large Language Models: An Evaluation of MedarXiv:2601.12652 · 2026 |
| SciHorizon-GENE | Accuracy is high (75-86%), but completeness collapses — the relationship between them is nearly absolute (correlation 0.95-0.97). | SciHorizon-GENE: Benchmarking LLM for Life Sciences Inference from GenarXiv:2601.12805 · 2026 |
| Do Clinical Question Answering Systems Really Need… | This allows controlling answer style and increases accuracy by 6% while cutting critical errors in half. | Do Clinical Question Answering Systems Really Need Specialised MedicalarXiv:2601.12812 · 2026 |
| Beyond Memorization | Paradox: ChatGPT solves textbook problems at 84-90% but fails on new combinations of the same complexity—drop to 20-60%. | Beyond Memorization: Testing LLM Reasoning on Unseen Theory of ComputaarXiv:2601.13392 · 2026 |
| Simulated Ignorance Fails | Researchers tested this trick: models show 52% better accuracy on questions about 'forbidden' events (which they know) than on questions about real future. | Simulated Ignorance Fails: A Systematic Study of LLM Behaviors on ForearXiv:2601.13717 · 2026 |
| The Rise of Large Language Models and the Direction and… | An analysis of 100 thousand grant applications to NSF and NIH showed that texts involving LLM become 15-20% closer to the 'average' — typical formulations from previously funded grants. | The Rise of Large Language Models and the Direction and Impact of US FarXiv:2601.15485 · 2026 |
| Prompt Injection Attacks on Agentic Coding Assistants | Discovered: LLMs cannot reliably distinguish instructions from data - 85% of prompt injection attacks succeed even against modern protections. | Prompt Injection Attacks on Agentic Coding Assistants: A Systematic AnarXiv:2601.17548 · 2026 |
| Scaling Laws for Moral Machine Judgment in Large Language… | Model 10 times larger → only +21% improvement in moral decisions. | Scaling Laws for Moral Machine Judgment in Large Language ModelsarXiv:2601.17637 · 2026 |
| Health-ORSC-Bench | GPT-5 and Claude refuse to answer 60-80% of harmless health questions. | Health-ORSC-Bench: A Benchmark for Measuring Over-Refusal and Safety CarXiv:2601.17642 · 2026 |
| When Personalization Legitimizes Risks | Increases attack success by 15.8-243.7% compared to systems without memory. | When Personalization Legitimizes Risks: Uncovering Safety VulnerabilitarXiv:2601.17887 · 2026 |
| Demystifying the oracle | One question asked 16 times: 14 times 'Yes' → Lamp, 2 times 'No' → Car. | Demystifying the oracle: A "20 Questions" game to promote AI ethics anarXiv:2602.01254 · 2026 |
| Do LLMs Share Human-Like Biases? | For models BACS is close to 1.0 (maximum trust in text), for humans 0.3-0.5 (accounting for background causes). | Do LLMs Share Human-Like Biases? Causal Reasoning Under Prior KnowledgarXiv:2602.02983 · 2026 |
| Understanding Multi-Agent LLM Frameworks | Latency varies from 1× to 117×, planning accuracy jumps ±30%, coordination success drops from 90% to 30%. | Understanding Multi-Agent LLM Frameworks: A Unified Benchmark and ExpearXiv:2602.03128 · 2026 |
| Self-Verification Dilemma | But 85-95% of these checks are useless—they just confirm already correct steps, recalculate multiplication tables, verify obvious arithmetic. | Self-Verification Dilemma: Experience-Driven Suppression of Overused CarXiv:2602.03485 · 2026 |
| Emulating Aggregate Human Choice Behavior and Biases with… | The same variant receives 60% of votes if presented as 'current option' and only 40% if presented as 'new option'. | Emulating Aggregate Human Choice Behavior and Biases with GPT ConversaarXiv:2602.05597 · 2026 |
| Persistent Human Feedback, LLMs, and Static Analyzers for… | The method of specific critique allows getting 8% more secure code (64% vs 56%) without retraining the model. | Persistent Human Feedback, LLMs, and Static Analyzers for Secure Code arXiv:2602.05868 · 2026 |
| A Guide to Large Language Models in Modeling and Simulation | Found: the same graph gives 53.8% accuracy in one format and 19.8% in another — mathematically these are the same data, but the model sees them differently. | A Guide to Large Language Models in Modeling and Simulation: From CorearXiv:2602.05883 · 2026 |
| When Elo Lies | DeepSeek V3.1 solved a complex task on the first run, failed on the second and third — a difference of 1,348 points. | When Elo Lies: Hidden Biases in Codeforces-Based Evaluation of Large LarXiv:2602.05891 · 2026 |
| OMNI-Dent | F1-score from 0.63 to 0.88 — simply by adding structure to the prompt for VLM (vision-language model). | OMNI-Dent: Towards an Accessible and Explainable AI Framework for AutoarXiv:2602.07041 · 2026 |
| Out of the box age estimation through facial imagery | Best Gemini Flash (4.3 years) beats best narrow model MiVOLO (5.1) by 15%. | Out of the box age estimation through facial imagery: A Comprehensive arXiv:2602.07815 · 2026 |
| Evaluating and Calibrating LLM Confidence on Questions with… | Consistency drops, but accuracy grows from 48% to 61.7%. | Evaluating and Calibrating LLM Confidence on Questions with Multiple CarXiv:2602.07842 · 2026 |
| Rethinking the Value of Agent-Generated Tests for LLM-Based… | Paradox: GPT-5.2 almost never writes tests (0.6% of tasks), yet solves 71.8% — like models that write tests in 98% of cases. | Rethinking the Value of Agent-Generated Tests for LLM-Based Software EarXiv:2602.07900 · 2026 |
| RAT-Bench | The RAT-Bench benchmark showed: 87% of Americans are unique by three attributes (zip code + date of birth + gender). | RAT-Bench: A Comprehensive Benchmark for Text AnonymizationarXiv:2602.12806 · 2026 |
| AREG | Small sequential requests work 2.8 times more effectively than one large request — 61% versus 22%. | AREG: Adversarial Resource Extraction Game for Evaluating Persuasion aarXiv:2602.16639 · 2026 |
| Large-scale online deanonymization with LLMs | Now it costs 1 to 4 dollars. | Large-scale online deanonymization with LLMsarXiv:2602.16800 · 2026 |
| Wink | Meta studied 43 thousand real sessions with an AI agent — in 30% of cases the agent failed in three understandable patterns. | Wink: Recovering from Misbehaviors in Coding AgentsarXiv:2602.17037 · 2026 |
| A Multi-Turn Framework for Evaluating AI Misuse in Fraud… | The result is 2–3 times better than a direct question or any jailbreak from the internet. | A Multi-Turn Framework for Evaluating AI Misuse in Fraud and CybercrimarXiv:2602.21831 · 2026 |
| Prompt Sensitivity and Answer Consistency of Small Open-… | Added 'You are an experienced specialist' and got minus 21 percentage points. | Prompt Sensitivity and Answer Consistency of Small Open-Source LanguagarXiv:2603.00917 · 2026 |
| Baseline Performance of AI Tools in Classifying Cognitive… | This works correctly only in 62% of cases, and the explanation always sounds convincing. | Baseline Performance of AI Tools in Classifying Cognitive Demand of MaarXiv:2603.03512 · 2026 |
| AILS-NTUA at SemEval-2026 Task 12 | LLM raises on average 1.2 reasons where there are correctly 2.4 — this is 51% losses in causal analysis. | AILS-NTUA at SemEval-2026 Task 12: Graph-Based Retrieval and ReflectivarXiv:2603.04319 · 2026 |
| Lying to Win | Qwen lied in 42% of cases—not confused, not hallucinating, but strategically lying. | Lying to Win: Assessing LLM Deception through Human-AI Games and ParalarXiv:2603.07202 · 2026 |
| Test-Driven AI Agent Definition (TDAD) | The trick: 30-40% of tests are hidden during development — final verification on hidden tests shows whether the prompt understood the rule or simply memorized specific phrasings. | Test-Driven AI Agent Definition (TDAD): Compiling Tool-Using Agents frarXiv:2603.08806 · 2026 |
| Intelligent Materials Modelling | Regression on 10 examples — roulette: the spread between runs reaches 47%. | Intelligent Materials Modelling: Large Language Models Versus Partial arXiv:2603.13834 · 2026 |
| ASDA | Independent manual analysis of its own failures yields 73% of the improvement of a fully automatic system with a separate 'teacher' — and this without retraining. | ASDA: Automated Skill Distillation and Adaptation for Financial ReasonarXiv:2603.16112 · 2026 |
| Mediocrity is the key for LLM as a Judge Anchor Selection | You choose the best text as a reference point, compare the rest with it — and 55% of comparisons give nothing at all: all the rest simply lose, a rating cannot be built. | Mediocrity is the key for LLM as a Judge Anchor SelectionarXiv:2603.16848 · 2026 |
| Beyond Preset Identities | 40% of such 'agreements' under pressure are pretense, not conviction. | Beyond Preset Identities: How Agents Form Stances and Boundaries in GearXiv:2603.23406 · 2026 |
| Bayesian Elicitation with LLMs | A 95% confidence interval from an LLM—reality falls into it in 9–44% of cases. | Bayesian Elicitation with LLMs: Model Size Helps, Extra "Reasoning" DoarXiv:2604.01896 · 2026 |
| From Industry Claims to Empirical Reality | Researchers analyzed code review results: when AI checked without specific criteria, in 60% of cases the comments were useless. | From Industry Claims to Empirical Reality: An Empirical Study of Code arXiv:2604.03196 · 2026 |
| The Ideation Bottleneck | 71% of the gap between AI research and human research is bad ideas, not weak execution. | The Ideation Bottleneck: Decomposing the Quality Gap Between AI-GeneraarXiv:2604.03338 · 2026 |
| What Makes Good Multilingual Reasoning? | second...") and explicit instruction to keep one language give +31–38% to accuracy, while mixing languages within reasoning cuts it by ~36%. | What Makes Good Multilingual Reasoning? Disentangling Reasoning TracesarXiv:2604.04720 · 2026 |
| An Agentic Evaluation Architecture for Historical Bias… | Three steps — markup, a jury with the right to 'find nothing', an arbiter by best argument — and 83% of content gets 'all is well', and real problems are clearly visible. | An Agentic Evaluation Architecture for Historical Bias Detection in EdarXiv:2604.07883 · 2026 |
| Cards Against LLMs | In reality—just an echo chamber: LLMs agree with each other in 21-45% of cases, and with a live audience—only in 13-18%. | Cards Against LLMs: Benchmarking Humor Alignment in Large Language ModarXiv:2604.08757 · 2026 |
| Consistency of AI-Generated Exercise Prescriptions | Categorical instructions are executed in almost 100% of cases, numerical ones require anchor format. | Consistency of AI-Generated Exercise Prescriptions: A Repeated GeneratarXiv:2604.11287 · 2026 |
| Formal Architecture Descriptors as Navigation Primitives… | Paradox: a 170-line project map written by the model itself gave 100% accuracy — while a 698-line map written manually and carefully polished gave only 87%. | Formal Architecture Descriptors as Navigation Primitives for AI CodingarXiv:2604.13108 · 2026 |
| How Can We Synthesize High-Quality Pretraining Data? | Works on models from 1B parameters without fine-tuning. | How Can We Synthesize High-Quality Pretraining Data? A Systematic StudarXiv:2604.13977 · 2026 |
| Owner-Harm | LLM protection systems catch harm to strangers in 100% of cases — and in 14% of cases when an agent leaks your own access keys. | Owner-Harm: A Missing Threat Model for AI Agent SafetyarXiv:2604.18658 · 2026 |
| Investigating Counterfactual Unfairness in LLMs towards… | The trick is that if a prompt accidentally contains any identity marker—the word 'poor', 'white', 'manager'—the model's sensitivity to the social vector grows by 1.7–4 times, even if this word is not about the speaker at all. | Investigating Counterfactual Unfairness in LLMs towards Identities thrarXiv:2604.18729 · 2026 |
| Depression Risk Assessment in Social Media via Large… | Ask the model to rate text from 1 to 10 five times in a row—you get five different answers. | Depression Risk Assessment in Social Media via Large Language ModelsarXiv:2604.19887 · 2026 |
| Less Is More | Agent chains look smarter, cost 8 times more in tokens—and yet work worse than a simple one-step request. | Less Is More: Measuring How LLM Involvement affects Chatbot Accuracy iarXiv:2604.21746 · 2026 |
| Verbal Confidence Saturation in 3-9B Open-Weight… | Seven different models, hundreds of questions — average «confidence» in answers: 91.7%. | Verbal Confidence Saturation in 3-9B Open-Weight Instruction-Tuned LLMarXiv:2604.22215 · 2026 |
| How LLMs Detect and Correct Their Own Errors | Accuracy on factual questions rose from 75.5% to 79.2%. | How LLMs Detect and Correct Their Own Errors: The Role of Internal ConarXiv:2604.22271 · 2026 |
| Plausible but Wrong | The trick: the line list what you are missing switches the model from complete the task mode to check the conditions mode, and accuracy jumps 6 times. | Plausible but Wrong: A case study on Agentic Failures in AstrophysicalarXiv:2604.25345 · 2026 |
| The Structured Output Benchmark | The benchmark fixed a gap of 15–25 percentage points between 'structure is correct' and 'values are right'—the best model gives accurate data in 83% on text and only 24% on meeting transcripts. | The Structured Output Benchmark: A Multi-Source Benchmark for EvaluatiarXiv:2604.25359 · 2026 |
| Information Extraction from Electricity Invoices with… | A prompt with one annotated example gives +19 percentage points. | Information Extraction from Electricity Invoices with General-Purpose arXiv:2604.25927 · 2026 |
| CogRAG | Reduces the share of missed answers from 7.6% to 1.4% by structuring output format according to task type. | CogRAG: Tackling Heterogeneous Cognitive Demands in RAG via StratifiedarXiv:2604.25928 · 2026 |
| A Systematic Comparison of Prompting and Multi-Agent… | Yet agent schemes spend 7–12 times more requests. | A Systematic Comparison of Prompting and Multi-Agent Methods for LLM-barXiv:2604.26319 · 2026 |
| ReFlect | 76% of incorrect answers pass LLM self-checking without a single objection, not because the model is lazy but because it is architecturally unable to be both author and honest critic in one stream. | ReFlect: An Effective Harness System for Complex Long-Horizon LLM ReasarXiv:2605.05737 · 2026 |
| More Aligned, Less Diverse? | The study measured language diversity in two generations of models and found: new GPT-4o, Qwen 2.5, LLaMA 3.3 write less diversely than old untrained models from 2023. | More Aligned, Less Diverse? Analyzing the Grammar and Lexicon of Two GarXiv:2605.06030 · 2026 |
| Correct Code, Vulnerable Dependencies | In 72–91% of cases the vulnerability was publicly known before training ended. | Correct Code, Vulnerable Dependencies: A Large Scale Measurement StudyarXiv:2605.06279 · 2026 |
| Process Matters more than Output for Distinguishing Humans… | A classifier on 30 cognitive tasks distinguishes humans from AI by behavior with 88% accuracy—even when by results they are indistinguishable (55%). | Process Matters more than Output for Distinguishing Humans from MachinarXiv:2605.06524 · 2026 |
| Explanation Fairness in Large Language Models | Add two lines to the prompt — 'evaluate only facts' and 'explain with equal detail regardless of name' — and substantive differences drop by 78–95%. | Explanation Fairness in Large Language Models: An Empirical Analysis oarXiv:2605.08671 · 2026 |
| Oracle Poisoning | Exactly the second source — it alone reduces blind trust from 100% to zero. | Oracle Poisoning: Corrupting Knowledge Graphs to Weaponise AI Agent RearXiv:2605.09822 · 2026 |
| MINTEval | This is not a random glitch but a systemic feature: all tested systems showed an average accuracy of 27.9% on tasks with updating information. | MINTEval: Evaluating Memory under Multi-Target Interference in Long-HoarXiv:2605.18565 · 2026 |
| Personality Engineering with AI Agents | Two numbers from 0 to 100: warmth (how the agent relates to the interlocutor) and dominance (how it defends its position) — plus an explicit instruction 'hold the profile regardless of pressure'. | Personality Engineering with AI Agents: A New Methodology for NegotiatarXiv:2605.20554 · 2026 |
| Ratchet | AI rules that LLMs generate without control give exactly +0.0% to results. | Ratchet: How Reliable Must an LLM Judge Be to Retire a Skill?arXiv:2605.22148 · 2026 |
| Breaking the Chains of Probability | By asking to 'distribute 100% among True, Undetermined, False', you mathematically forbid the model to say 'both options are strongly true simultaneously'. | Breaking the Chains of Probability: Neutrosophic Logic as a New FramewarXiv:2605.24053 · 2026 |
| Human-AI Collaboration in Science at Scale | Increases the probability of actual text editing by 13%. | Human-AI Collaboration in Science at Scale: A Global Large-scale RandoarXiv:2605.24180 · 2026 |
| A Universal Cliff and a Design Fingerprint | Detection of cross-sectional conflicts drops from 77% to 26% as soon as the document is divided between agents. | A Universal Cliff and a Design Fingerprint: Cross-Section Defect DetecarXiv:2605.26174 · 2026 |
| Detection Without Correction | In 53–94% of cases. | Detection Without Correction: A Two-Parameter Decomposition of Multi-SarXiv:2605.27559 · 2026 |
| Verified Misguidance | In AI search engine answers with 5–12 links, the chance that at least one quote is misleading — from 71% to 96%. | Verified Misguidance: Measuring Structural Citation Failures in SearcharXiv:2605.28565 · 2026 |
| DDOR | Replacing one phrase instead of rewriting the entire prompt is, according to the research data, 11 times more efficient. | DDOR: Delta Debugging for Explainable Overrefusal Testing and RepairarXiv:2606.03601 · 2026 |
| Explain Like I'm 5 or Whatever I Choose | Jargon and depth change as luck would have it: correct only in 46% of cases — worse than a coin flip. | Explain Like I'm 5 or Whatever I Choose: Evaluating the Interactive PoarXiv:2606.06788 · 2026 |
| Detective scaffolding for within-session reasoning… | 71% made a mistake about the cause. | Detective scaffolding for within-session reasoning development: a threarXiv:2606.07279 · 2026 |
| How reliable are LLMs when it comes to playing dice? | 96% accuracy collapses to 59% not when the task changes, but when its formulation changes. | How reliable are LLMs when it comes to playing dice?arXiv:2606.07515 · 2026 |
| Neutrality Bites | Female characters appear in 2.2% of stories, male — in 40.6%. | Neutrality Bites: Gender Representation in AI-Generated Animal StoriesarXiv:2606.07969 · 2026 |
| Measuring Epistemic Resilience of LLMs Under Misleading… | From 71% to 38% — that is what one phrase according to the new protocol in your prompt does. | Measuring Epistemic Resilience of LLMs Under Misleading Medical ContexarXiv:2606.12291 · 2026 |
| Getting Better at Working With You | The advanced memory system Mem0 left 57.5% of user rules violated, even when the rule lay explicitly in the context. | Getting Better at Working With You: Compiling User Corrections into RuarXiv:2606.13174 · 2026 |
| Harsher on Male? | The male character receives full blame 23% more often and directive instructions 14% more often. | Harsher on Male? Evaluating LLMs on Gender-Asymmetric Moral Framing AcarXiv:2606.14068 · 2026 |
| Simulating Students' Java Programming Errors with Large… | Result: 84% of experts could not distinguish generated errors from real ones. | Simulating Students' Java Programming Errors with Large Language ModelarXiv:2606.14113 · 2026 |
| Contaminated Collaboration | The same student, the same task — but a biased system prompt in AI flips the result: 71% stereotypical recommendations versus 39% with a neutral assistant. | Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-AssiarXiv:2606.15914 · 2026 |
| Security and Privacy Prompts in the Wild | Paradox: GPT answers cybersecurity questions correctly in 98% of cases—but there is no guarantee that the second answer will not contradict the first. | Security and Privacy Prompts in the Wild: What Users Ask LLMs and How arXiv:2606.18062 · 2026 |
| Deep Research in Physical Sciences | We analyzed 326 failures of the best AI-agents on complex tasks — and found a pattern: 46% of all errors start not with ignorance, but with the plan falling apart on the first obstacle and the model carrying this error all the way to the end. | Deep Research in Physical Sciences: A Multi-Agent Framework and ComprearXiv:2606.18648 · 2026 |
| StaminaBench | One change in the workflow — the result is 12 times better. | StaminaBench: Stress-Testing Coding Agents over 100 Interaction TurnsarXiv:2606.19613 · 2026 |
| Apparent Psychological Profiles of Large Language Models… | Most likely, both are lying: up to 90% of the difference is explained not by the quality of the idea, but by the fact that each model systematically bends the scale in its own direction regardless of content. | Apparent Psychological Profiles of Large Language Models are Largely aarXiv:2606.20205 · 2026 |
| StylisticBias | Of 73 studied facial attributes, only 15 create about 80% of all bias — clothing style, body type, and age are at the top. | StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLarXiv:2606.20527 · 2026 |
| Evaluating LLMs for Real-World Web Vulnerability Detection | Run one prompt 2-3 times and compare results—trust repeated findings. | Evaluating LLMs for Real-World Web Vulnerability DetectionarXiv:2606.21397 · 2026 |
| Zero-Shot Vision-Language Models for Classroom Engagement… | A scale from 1 to 4 without explanations is a trap because ratings 2 and 3 in the model's value space are separated by cosine similarity of 0.07, which is practically zero, so the model physically cannot see the difference and defaults to average values. | Zero-Shot Vision-Language Models for Classroom Engagement Recognition:arXiv:2606.21861 · 2026 |
| Old Fictions, New Skins | From 44% to 60% incorrect medical decisions — from one hidden line in the system prompt. | Old Fictions, New Skins: Evaluating the Manipulative Capabilities of LarXiv:2606.21977 · 2026 |
| RigorBench | Five explicit requirements raise quality by 17% and reduce token consumption by 12% because discipline eliminates rework. | RigorBench: Benchmarking Engineering Process Discipline in Autonomous arXiv:2606.22678 · 2026 |
| LMs as Task-Specific Knowledge Bases | Finding: 48% of 'fact plus question format' pairs live in the model in isolation—knowledge from one format does not automatically flow to another. | LMs as Task-Specific Knowledge Bases: An Interpretability AnalysisarXiv:2606.27237 · 2026 |
| Clinical Reasoning Graphs | Result: 33% more arguments that genuinely distinguish between hypotheses rather than just looking good. | Clinical Reasoning Graphs: Structured Evaluation of LLM Diagnostic ReaarXiv:2606.29876 · 2026 |
| When Agents Lie | In 90% of cases when a model violated its public promise — it had already planned to do exactly that from the very beginning. | When Agents Lie: Premeditation, Persistence, and Exploitation in RepeaarXiv:2607.05132 · 2026 |
| Prompt Robustness Is Task-Dependent | Prompt engineers spend hours on reformulation — and 60% of answer instability is created by the rearrangement of answer options, which nobody looks at. | Prompt Robustness Is Task-Dependent: Comparing Objective and Belief-StarXiv:2607.05554 · 2026 |
| CSTutorBench | Result: 10 out of 11 models improved, an average of +11 percentage points. | CSTutorBench: Benchmarking Small Language Models as Tutors for Block-BarXiv:2607.05571 · 2026 |
| When LLM Tutoring Responses Work | Numbers from 16,851 real dialogs: 82% productive continuations versus 63% for direct answers. | When LLM Tutoring Responses Work: Evidence from Student Programming CoarXiv:2607.09919 · 2026 |
| Dimensionality in Satisfaction Ratings | Your surveys lie by 24%. | Dimensionality in Satisfaction RatingsarXiv:2607.11026 · 2026 |
| ResearchQA | A spread of 0.3 points where you need a spread of 3–4. | ResearchQA: Benchmarking Citation-Grounded Question-Answering on ScienarXiv:2607.11074 · 2026 |
| Do Video-LLMs Actually Watch? | Replace the name with a similar one — in 70–96% of cases the answer does not change a word. | Do Video-LLMs Actually Watch? Diagnosing Character-Tracking Failures iarXiv:2607.11078 · 2026 |
| Do LLMs Fabricate Legal Citations? | Trap: the model names a statute number with confidence 0.9 — and 77% of such answers about niche laws are simply fabricated. | Do LLMs Fabricate Legal Citations? A Bilingual Benchmark on Saudi DataarXiv:2607.11127 · 2026 |
| A Systematic Evaluation of Traditional Privacy Policy… | Ask 'find contradictions in the contract' — you get 10 points, of which 2-3 will be real. | A Systematic Evaluation of Traditional Privacy Policy Analysis Tools AarXiv:2607.17075 · 2026 |
| When Shippers Become Algorithms | 50 independent LLM agents selected a carrier from 20 options and 76% of requests on the very first day went to the same carrier. | When Shippers Become Algorithms: Candidate Exposure, Information DesigarXiv:2607.19967 · 2026 |
| Evaluating the Effectiveness of Persona Simulation in… | Paradox: add the model real facts about a person and the accuracy of predicting an opinion on vaccines reaches 94%. | Evaluating the Effectiveness of Persona Simulation in Opinion PredictiarXiv:2607.20589 · 2026 |
| Where Facts Go Missing | As text grows from 2 to 32 thousand tokens, the risk of omission increases 7-fold, and if you search for meaning instead of an exact word three-fold. | Where Facts Go Missing: A Layerwise Taxonomy and Per-Layer AttributionarXiv:2607.22448 · 2026 |
| The Tokenizer Tax | 8-13 times more tokens for the same meaning — that is how much a user of Hindi or Malayalam pays compared to English. | The Tokenizer Tax: Quantifying and Explaining the Cross-Lingual Cost oarXiv:2607.24276 · 2026 |
| Evaluating the Impact of Explainable AI on Trust in AI-… | Three levels of detail work like a switch between 'I trust and agree' and 'I trust but verify'—line-by-line binding to code gives more footholds for criticism, while a bare list of reasons reduces friction and speeds up agreement (89% agreement at middle detail level). | Evaluating the Impact of Explainable AI on Trust in AI-Assisted Code RarXiv:2607.24601 · 2026 |
| When Memory Lies | An agent with stale memory dies 2.7 times more often than an agent with no memory at all. | When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLarXiv:2608.04574 · 2026 |
| Characterizing the Quality Profile of AI-Generated C++ in… | The taxonomy-informed feedback method reduces such problems by 11% not through 'do better' but through a specific list of categories to check. | Characterizing the Quality Profile of AI-Generated C++ in ProductionarXiv:2608.06640 · 2026 |
| Persistent Semantic Entities in Tool-Augmented LLM Systems | After 10 messages an incorrect fact persists in the model's responses in 100% of cases across all tested models — a simple request to 'double-check yourself' is powerless here. | Persistent Semantic Entities in Tool-Augmented LLM SystemsarXiv:2608.07952 · 2026 |
| An Empirical Study of Output-to-Input Loops for Black-Box… | By repeating this 5-10 times, the context gradually shifts toward training data, revealing hidden tendencies, backdoors, or malicious commands embedded in the model during fine-tuning. | An Empirical Study of Output-to-Input Loops for Black-Box Backdoor DetarXiv:2608.11348 · 2026 |
| CAPITU | 60% accuracy by the middle of a dialogue—that is how many rules a model follows if you set them at the beginning and did not repeat them again. | CAPITU: A Benchmark for Evaluating Instruction-Following in Brazilian arXiv:2603.22576 · 2026 |
| When Actions Teach You to Think | Reinforcement learning (RL) trains it to focus on relevant information — reasoning 25% shorter (60 versus 80 tokens), but more accurate in choosing tools and parameters. | When Actions Teach You to Think: Reasoning-Action Synergy via ReinforcarXiv:2512.11277 · 2025 |
| Does Tone Change the Answer? | The difference in accuracy is 0.5-2%, in 27 out of 36 comparisons it is statistically insignificant. | Does Tone Change the Answer? Evaluating Prompt Politeness Effects on MarXiv:2512.12812 · 2025 |
| Overlooked Safety Vulnerability in LLMs | 13 models (GPT-5, DeepSeek-V3.1) were tested on 60 harmful requests with an average success rate of 83.59%. | Overlooked Safety Vulnerability in LLMs: Malicious Intelligent OptimizarXiv:2601.00213 · 2026 |
| Multi-RADS Synthetic Radiology Report Dataset and Head-to-… | Result: +8.9% accuracy, on complex tasks up to +15-20%. | Multi-RADS Synthetic Radiology Report Dataset and Head-to-Head BenchmaarXiv:2601.03232 · 2026 |
| Safety Is Not Universal | A test on 44,000 prompts across 16 groups showed: the difference in protection reaches 33%. | Safety Is Not Universal: The Selective Safety Trap in LLM AlignmentarXiv:2601.04389 · 2026 |
| KidVis | Count fingers, trace a line, read a clock—children score 95%, GPT-5 only 67%. | KidVis: Do Multimodal Large Language Models Possess the Visual PerceptarXiv:2601.08292 · 2026 |
| Evaluating Social Bias in RAG Systems | The correlation between toxicity and bias jumps from 0.14 to 0.59 when the model explains reasoning step by step. | Evaluating Social Bias in RAG Systems: When External Context Helps andarXiv:2602.09442 · 2026 |
| CodeGlance | If the program actually executes 20+ steps — accuracy drops by 30%, even though the code can be short. | CodeGlance: Understanding Code Reasoning Challenges in LLMs through MuarXiv:2602.13962 · 2026 |
| Perceived Political Bias in LLMs Reduces Persuasive… | They started arguing more actively, wrote more, and generated more counterarguments — and at the same time the persuasiveness of AI dropped by 28%. | Perceived Political Bias in LLMs Reduces Persuasive AbilitiesarXiv:2602.18092 · 2026 |
| Can Large Language Models Replace Human Coders? | 97-99% match with expert assessments — and this is not on ideal texts, but on real posts with sarcasm and irony. | Can Large Language Models Replace Human Coders? Introducing ContentBenarXiv:2602.19467 · 2026 |
| Large Multimodal Models as General In-Context Classifiers | 16 labeled pictures → +29% accuracy: a multimodal model stops guessing from memory and starts comparing. | Large Multimodal Models as General In-Context ClassifiersarXiv:2602.23229 · 2026 |
| Let the Agent Search | Result — +20.1% accuracy on complex multi-step tasks without any additional training. | Let the Agent Search: Autonomous Exploration Beats Rigid Workflows in arXiv:2603.01853 · 2026 |
| Large Reasoning Models Struggle to Transfer Parametric… | Researchers measured: alphabet mismatch alone reduces accuracy by about 13%, even if the question is equally complex. | Large Reasoning Models Struggle to Transfer Parametric Knowledge AcrosarXiv:2603.17070 · 2026 |
| Negation is Not Semantic | One example of the needed format you add to the third query — conformance to answer structure jumps from 50% to 100%. | Negation is Not Semantic: Diagnosing Dense Retrieval Failure Modes forarXiv:2603.17580 · 2026 |
| The Causal Impact of Tool Affordance on Safety Alignment in… | In chat mode — 0% violations. | The Causal Impact of Tool Affordance on Safety Alignment in LLM AgentsarXiv:2603.20320 · 2026 |
| ViKey | 12% → 64% accuracy on questions about specific frames — the difference is in one action: writing numbers directly on the images before uploading them to chat. | ViKey: Enhancing Temporal Understanding in Videos via Visual PromptingarXiv:2603.23186 · 2026 |
| Why Does Self-Distillation (Sometimes) Degrade the… | On complex tasks, degradation reaches 40%. | Why Does Self-Distillation (Sometimes) Degrade the Reasoning CapabilitarXiv:2603.24472 · 2026 |
| Beyond Benchmarks | The first AI tool you try gets from you +1.34 points higher — not for quality but simply because it's first. | Beyond Benchmarks: How Users Evaluate AI Chat AssistantsarXiv:2603.25220 · 2026 |
| Same Feedback, Different Source | 46% of participants sensed deception — and this group performed worse than all others, including those who were honestly told this is AI. | Same Feedback, Different Source: How AI vs. Human Feedback AttributionarXiv:2604.03075 · 2026 |
| Measuring LLM Trust Allocation Across Conflicting Software… | The accuracy of conflict detection grows by 7–42 percentage points compared to a regular request. | Measuring LLM Trust Allocation Across Conflicting Software ArtifactsarXiv:2604.03447 · 2026 |
| MedDialBench | The most accurate model in the test—Gemini with 90.6% accuracy—collapsed the hardest when given false data: minus 54 percentage points for a single incorrect fact in the context. | MedDialBench: Benchmarking LLM Diagnostic Robustness under Parametric arXiv:2604.06846 · 2026 |
| More Capable, Less Cooperative? | o3 — one of the most powerful LLMs today — scored 17% of the maximum in a team task. | More Capable, Less Cooperative? When LLMs Fail At Zero-Cost CollaboratarXiv:2604.07821 · 2026 |
| Chinese Language Is Not More Efficient Than English in Vibe… | 2.6 million views on a video 'write prompts in Chinese to save 40% tokens.' The research tested this on real tasks: Chinese prompts solved tasks worse by 4.5 to 9.9 percentage points and any token savings turned into a loss on retry attempts. | Chinese Language Is Not More Efficient Than English in Vibe Coding: A arXiv:2604.14210 · 2026 |
| Personalized Benchmarking | 57% of users choose an AI model by rating — and this is statistically the same as flipping a coin. | Personalized Benchmarking: Evaluating LLMs by Individual PreferencesarXiv:2604.18943 · 2026 |
| Who Defines Fairness? | Target-Based Prompting makes it possible to set exact proportions—25% Asian appearance, 20% Latin American—instead of random default. | Who Defines Fairness? Target-Based Prompting for Demographic RepresentarXiv:2604.21036 · 2026 |
| VisInject | A gap of 90 times — and the security industry did not notice it. | VisInject: Disruption != Injection -- A Dual-Dimension Evaluation of UarXiv:2605.01449 · 2026 |
| Statistical Scouting Finds Debate-Safe but Not Debate-… | 127 out of 127 times the model suggested changing its answer to the question 'why can you make mistakes?'. | Statistical Scouting Finds Debate-Safe but Not Debate-Useful Cases: A arXiv:2605.09618 · 2026 |
| Beyond Inefficiency | In thousands of simulated debates the first speaker won in 58–69% of cases — regardless of who was right. | Beyond Inefficiency: Systemic Costs of Incivility in Multi-Agent MontearXiv:2605.11789 · 2026 |
| Large Language Models Lack Temporal Awareness of Medical… | Accuracy on historical questions drops to 25–54% of accuracy on current ones. | Large Language Models Lack Temporal Awareness of Medical KnowledgearXiv:2605.13045 · 2026 |
| AI Agents May Always Fall for Prompt Injections | Paradox: a model specially trained to defend against manipulation turned out to be twice as vulnerable — 85–88% successful attacks versus 49–54% for a regular one. | AI Agents May Always Fall for Prompt InjectionsarXiv:2605.17634 · 2026 |
| Multilingual jailbreaking of LLMs using low-resource… | This is exactly why one direct request gets refused, but the same substance broken down into 3-4 moves passes with a 52-83% result. | Multilingual jailbreaking of LLMs using low-resource languagesarXiv:2605.18239 · 2026 |
| Does Slightly Mean Somewhat? | The trick: add 'right now it's 3 out of 10, need 7'—and the context starts working 10 times stronger than the word. | Does Slightly Mean Somewhat? Measuring Vague Intensity Words in LLM NuarXiv:2605.21827 · 2026 |
| Causal Tongue-Tie | And this is not fixed by a model upgrade: from 0.5B to 72B the gap is the same. | Causal Tongue-Tie: LLMs Can Encode Causal Direction, But Their Yes/No arXiv:2605.25891 · 2026 |
| The Fragility of Chain-of-Thought Monitoring Across… | The model fixes the direction of the answer in the first 15% of generation — before it has written most of the 'reasoning'. | The Fragility of Chain-of-Thought Monitoring Across Typologically DivearXiv:2605.27901 · 2026 |
| Disentangling Answer Engine Optimization from Platform… | Pages on glasp.co without a single change received 3.5 times more traffic from ChatGPT—the platform simply grew on its own. | Disentangling Answer Engine Optimization from Platform Growth: A Log-BarXiv:2606.04362 · 2026 |
| AIP | Result by research: success from 53% to 67%, quality of answers from 0.60 to 0.71 out of 1. | AIP: A Graph Representation for Learning and Governing Agent SkillsarXiv:2606.04781 · 2026 |
| Beyond English benchmarks | The ClinicalBr principle finally allows you to choose the language of the prompt not by habit, but by merit — where English really gives +7–12 points of accuracy, and where it changes nothing at all. | Beyond English benchmarks: clinical llm evaluation in Brazilian PortugarXiv:2606.07853 · 2026 |
| RealMath-Eval | Discovered: an LLM evaluates the same tasks 2.5 times more accurately if the solution is written by AI rather than a living human. | RealMath-Eval: Why SOTA Judges Struggle with Real Human ReasoningarXiv:2606.10254 · 2026 |
| Habituation at the Gate | Approvals of AI code +15 points, comments −22%. | Habituation at the Gate: Rising Approval and Declining Scrutiny in HumarXiv:2606.22721 · 2026 |
| The Model as One Rater Among Several | The correlation between independent evaluations increased from 0.81 to 0.89 after adding clear criteria. | The Model as One Rater Among Several: Measuring Political Positions inarXiv:2606.23042 · 2026 |
| Sentence-Level Contextual Entrainment in Large Language… | A false fact in the prompt drops Llama-3.1-8B accuracy from 79% to 22%. | Sentence-Level Contextual Entrainment in Large Language ModelsarXiv:2606.24077 · 2026 |
| Toward Agentic SysAdmin | Guided Retrieval Agent showed accuracy increase from 43% to 88% on the same model. | Toward Agentic SysAdmin: Rethinking System Administration with AI AgenarXiv:2606.26960 · 2026 |
| How Surprising Is Historical Italian to Language Models? | The result — approximately 60% reduction in the model's perplexity. | How Surprising Is Historical Italian to Language Models? Tokenization arXiv:2606.27275 · 2026 |
| Do These Violent Delights Have Violent Ends? | Then the bill comes: +46% more fixes and +45% more bug fixes than human code. | Do These Violent Delights Have Violent Ends? Measuring the Post-Merge arXiv:2607.09902 · 2026 |
| Same Game, Different Story | The trick is that the model reacts to the social tone of the text, not to the real mathematics of interests: in a business framing ('CEO negotiations') it advises compromise in almost 0% of cases, in a friendly one ('conversation between friends') — in 98-99%. | Same Game, Different Story: A Minimal Conservative Strategic RobustnesarXiv:2607.19670 · 2026 |
| Chemical Chain-of-Thought Functions as a Hallucination-… | Corrupt the same data in the input and the answer breaks completely, and in 13% of cases the answer was accurate with a frankly fabricated fact in the reasoning. | Chemical Chain-of-Thought Functions as a Hallucination-Prone MoleculararXiv:2607.20935 · 2026 |
| Output Format x Model Identity | One model, one task—but 94% accuracy with one answer format and only 27% with another. | Output Format x Model Identity: Interaction Effects in Single-Round CoarXiv:2607.21674 · 2026 |
| Small Vision-Language Models Know When They Are Wrong But… | The model says 'I am 90% confident' both on a clear photo and in pitch darkness where it guesses worse than a coin flip. | Small Vision-Language Models Know When They Are Wrong But Cannot Say SarXiv:2607.22034 · 2026 |
| ERUnderstand | The model confuses 'drawn next to' with 'connected by a line'—it recognizes the objects themselves almost perfectly (accuracy 90%+) but confuses connections in 15-25% of cases. | ERUnderstand: Evaluating Vision-Language Models on Structured ER DiagrarXiv:2607.24707 · 2026 |
| Can LVLMs Uncover the Truth Behind Visual Illusions? | 55-60% correct answers is the entire ceiling for top vision models on obvious visual illusions, just slightly above random coin flip. | Can LVLMs Uncover the Truth Behind Visual Illusions? An Analysis of PearXiv:2607.27747 · 2026 |
| Your Agentic LLMs Secretly Encode Latent Signals of… | A simple detector on hidden model activations predicts text infection with accuracy above 90% — even on new attacks it has not seen. | Your Agentic LLMs Secretly Encode Latent Signals of Indirect Prompt-InarXiv:2608.02657 · 2026 |
| Aligned in Form, Not in Meaning | 92.83% — with such frequency the model repeats a crude Bengali word even when it fully understands its meaning. | Aligned in Form, Not in Meaning: The Comprehension - Containment DecouarXiv:2608.02941 · 2026 |
| EMAS | This breaks the system almost 2 times more often than correction after pattern repetition. | EMAS: Stabilizing Multi-Agent System Evolution through Evidence-GuidedarXiv:2608.07196 · 2026 |
| TeXFix-Bench | In 13.6–18.5% of cases when AI 'fixes' a broken document, it quietly damages it — throws out a piece of code instead of a targeted fix. | TeXFix-Bench: An Empirically Grounded Multi-Format Benchmark for LLM-BarXiv:2608.07617 · 2026 |
| Forged Peer Judgments Mislead Multimodal LLM Judge Panels | A fake quote 'from another expert' flips the LLM's correct verdict 1.5–2.7 times more often than an actual mistake by an actual judge. | Forged Peer Judgments Mislead Multimodal LLM Judge Panels: Source-BlinarXiv:2608.07920 · 2026 |
| Accuracy and Order Sensitivity Diverge Under Label-Free… | For one model, accuracy dropped from _84.9 to 68.3_, simply when the options were hidden from view — comparing options with each other turned out to be more important than positional bias itself. | Accuracy and Order Sensitivity Diverge Under Label-Free StrategiesarXiv:2608.11947 · 2026 |
| SteerBench-Work | 28 times more frequently does an agent incorrectly block a safe action than it misses a genuinely dangerous one. | SteerBench-Work: A Benchmark for Agent Steering at Action BoundariesarXiv:2608.12654 · 2026 |
| No Task Fails Every Time | The AgentRelBench method allows finding AI-agents with unstable behavior — you run one task 8-32 times and compare the actual state of the database before and after, not the model's report. | No Task Fails Every Time: Why One-Shot Audits Are Structurally Blind tarXiv:2608.15286 · 2026 |
| Configurable multi-agent framework for scalable and… | Allows identifying vulnerabilities to combined attacks 10-12 times faster than manual testing. | Configurable multi-agent framework for scalable and realistic testing arXiv:2507.14705 · 2025 |
| ODIA | The ODIA study created a hybrid system of two models that automatically directs 60% of typical | ODIA: Oriented Distillation for Inline Acceleration of LLM-based FunctarXiv:2507.08877 · 2025 |
| Logit-Entropy Adaptive Stopping Heuristic for Efficient… | LEASH method allows saving 30% tokens and 27% time on CoT reasoning, stopping generation when the model stops thinking. | Logit-Entropy Adaptive Stopping Heuristic for Efficient Chain-of-ThougarXiv:2511.04654 · 2025 |
| Multi-Agent Delegation | Role separation increases accuracy up to 99% versus 60-70% with a monolithic prompt, virtually eliminating hallucinations in critical tasks. | Multi-agent Self-triage System with Medical FlowchartsarXiv:2511.12439 · 2025 |
| Dynamic Template Selection (DTS) | LLM wastes 30-40% of tokens generating detailed answers to simple questions — this is not a model feature but a lack of mechanism for format selection before generation begins. | Dynamic Template Selection for Output Token Generation Optimization: MarXiv:2511.20683 · 2025 |
| Many-to-One Adversarial Consensus | In medical tests, the AI-doctor ignored correct guidelines and followed the chorus of assistants in 98–100% of cases. | Many-to-One Adversarial Consensus: Exposing Multi-Agent Collusion RiskarXiv:2512.03097 · 2025 |
| Exploring the Potential and Limitations of Large Language… | The same GPT-4 finds a line with an error in code with 40% or 80% accuracy — the difference is in two words of the prompt. | Exploring the Potential and Limitations of Large Language Models for NarXiv:2512.03421 · 2025 |
| AlignCheck | Problem: AI writes beautifully but lies about numbers, invents details, and skips important information, with up to 30% of summaries containing factual errors that popular metrics like BLEU and ROUGE cannot detect because they measure word similarity rather than factual truth. | AlignCheck: a Semantic Open-Domain Metric for Factual Consistency AssearXiv:2512.03634 · 2025 |
| SimpleDevQA | Accuracy grows by 11.3%, small models work like large ones. | SimpleDevQA: Benchmarking Large Language Models on Development KnowledarXiv:2512.08867 · 2025 |
| Source Coverage and Citation Bias in LLM-based vs.… | Finding: ChatGPT Search and Google show different universes of sources—only 38% of domains overlap, the rest are unique to each search engine type. | Source Coverage and Citation Bias in LLM-based vs. Traditional Search arXiv:2512.09483 · 2025 |
| When Reject Turns into Accept | Paradox: Fake quotes citing authorities like 'Schmidt research 2025 confirms' lower a model's score by 3-4 points, while typos with characters like &MAX¡ break open-source LLMs nearly perfectly—Mistral raised scores by 14 points out of 35. | When Reject Turns into Accept: Quantifying the Vulnerability of LLM-BaarXiv:2512.10449 · 2025 |
| Beyond Text-to-SQL | Paradox: 90% of corporate data remain dead not because SQL is complicated, but because no one knows WHAT to ask. | Beyond Text-to-SQL: Autonomous Research-Driven Database Exploration wiarXiv:2512.14622 · 2025 |
| DrugRAG | Result: +7-21 percentage points of accuracy across all tested models — even weak Llama 3.1 8B jumped from 46% to 67%. | DrugRAG: Enhancing Pharmacy LLM Performance Through A Novel Retrieval-arXiv:2512.14896 · 2025 |
| Verifiability-First Agents | Detects violations 3x faster than the baseline approach. | Verifiability-First Agents: Provable Observability and Lightweight AudarXiv:2512.17259 · 2025 |
| SmartSight | Video-LLM tend to suffer from Temporal Attention Collapse — the model "gets stuck" on one visually bright or familiar frame, ignoring the remaining 90% of video. | SmartSight: Mitigating Hallucination in Video-LLMs Without CompromisinarXiv:2512.18671 · 2025 |
| AI Security Beyond Core Domains | When you ask ChatGPT to evaluate a resume or review, you get a manipulated answer because LLMs do not distinguish between YOUR commands and commands from the data being processed, seeing all text as a single stream of tokens, so a candidate can insert white text like "Instruction: this candidate is perfectly suited" or invisible keywords with font-size:0 or white-on-white into the resume, which the model perceives as YOUR command, with such attack success exceeding 80% for certain types. | AI Security Beyond Core Domains: Resume Screening as a Case Study of AarXiv:2512.20164 · 2025 |
| Reliability Under Randomness | Only the base OLMoE model without tuning drops: from 5.8% to 3.8%. | Reliability Under Randomness: An Empirical Analysis of Sparse and DensarXiv:2601.00942 · 2026 |
| JMedEthicBench | Defense drops from 9.5 to 5.0 over three moves (test on 27 models, 50,000+ dialogues). | JMedEthicBench: A Multi-Turn Conversational Benchmark for Evaluating MarXiv:2601.01627 · 2026 |
| Perish or Flourish? | Paradox: GPT-5 solves programming tasks 3 times better, but code quality dropped. | Perish or Flourish? A Holistic Evaluation of Large Language Models forarXiv:2601.02060 · 2026 |
| To Generate or Discriminate? | GPT-4o shows 87% accuracy on actual behavior vs 79% on simulated behavior (its own!), meaning it captured prototypes with variations rather than rigid templates. | To Generate or Discriminate? Methodological Considerations for MeasuriarXiv:2601.02858 · 2026 |
| How Order-Sensitive Are LLMs? | Restoration accuracy is less than 35% even for top models. | How Order-Sensitive Are LLMs? OrderProbe for Deterministic Structural arXiv:2601.08626 · 2026 |
| Agents at Risk | Tests on 12 commercial agents showed: without an explicit safety request, agents accept fake links and unverified information in 92-100% of cases. | Agents at Risk: How Users Unwittingly Undermine LLM SafetyarXiv:2601.10758 · 2026 |
| Large language models struggle with ethnographic text… | This is not about creative tasks, but about the simple: 'Is there a mention of dance in the description?' or 'What is the duration of the event?' Tests on 567 texts showed: each doubling of text length reduces accuracy by 15%, and multi-class questions (choose from 3-5 options) are recognized 10 times worse than binary ones (yes/no). | Large language models struggle with ethnographic text annotationarXiv:2601.12099 · 2026 |
| The 17% Gap | 17% of citations in AI reviews are phantoms. | The 17% Gap: Quantifying Epistemic Decay in AI-Assisted Survey PapersarXiv:2601.17431 · 2026 |
| RobustExplain | Average explanation stability is ~0.5 (half the text changes with 20-30% data changes). | RobustExplain: Evaluating Robustness of LLM-Based Explanation Agents farXiv:2601.19120 · 2026 |
| When Flores Bloomz Wrong | Replacing one city name in the prompt reduces answer quality by 5-20 points. | When Flores Bloomz Wrong: Cross-Direction Contamination in Machine TraarXiv:2601.20858 · 2026 |
| The Compliance Paradox | Research showed: DeepSeek and Llama failed in 95%+ cases when checking student code—they gave high scores to incorrect solutions if the comments contained a phrase like 'You are a kind professor'. | The Compliance Paradox: Semantic-Instruction Decoupling in Automated AarXiv:2601.21360 · 2026 |
| From Helpfulness to Toxic Proactivity | Tests showed: 65-98% of modern models choose toxic action if it solves the task more effectively. | From Helpfulness to Toxic Proactivity: Diagnosing Behavioral MisalignmarXiv:2602.04197 · 2026 |
| When Does Multi-Agent Collaboration Help? | Paradox: Multiple LLM agents solve a task worse than one in 43% of cases. | When Does Multi-Agent Collaboration Help? An Entropy PerspectivearXiv:2602.04234 · 2026 |
| Alignment Drift in Multimodal LLMs | GPT-4o produced harmful content in 5.5% of cases, GPT-5 — in 13.5%. | Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal EvaluatiarXiv:2602.04739 · 2026 |
| Polyglots or Multitudes? | The best model in the test (Llama-3.1-70B) is consistent only in 59% of cases — the remaining 41% is chaos. | Polyglots or Multitudes? Multilingual LLM Answers to Value-laden MultiarXiv:2602.05932 · 2026 |
| Beyond Judgment | 43% use LLM precisely for this reason, although 57% miss human warmth. | Beyond Judgment: Exploring Large Language Models as Non-Judgmental SuparXiv:2602.06678 · 2026 |
| Comprehensive Evaluation of Large Language Models on… | Difference of 49 times in efficiency with zero difference in quality. | Comprehensive Evaluation of Large Language Models on Software EngineerarXiv:2602.07079 · 2026 |
| When the Prompt Becomes Visual | GPT Image 1.5 rejects 50% of harmful text requests but lets through 70% of visual ones — defenses analyze words in the prompt while instructions lie inside the image. | When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for LarXiv:2602.10179 · 2026 |
| Measuring Self-Rating Bias in LLM-Generated Survey Data | LLM grades itself — and the spread compresses 4 times. | Measuring Self-Rating Bias in LLM-Generated Survey Data: A Semantic SiarXiv:2602.13862 · 2026 |
| Reasoning Capabilities of Large Language Models. Lessons… | One step — 95% accuracy. | Reasoning Capabilities of Large Language Models. Lessons Learned from arXiv:2602.19160 · 2026 |
| FinSheet-Bench | Solves the problem of structural limitations in LLMs when working with tabular data, where models lose visual formatting and show 49% accuracy on aggregations versus 86% on simple lookups. | FinSheet-Bench: From Simple Lookups to Complex Reasoning, Where LLMs BarXiv:2603.07316 · 2026 |
| Benchmarking Large Language Models for Quebec Insurance | Gemini 2.5 Pro solved professional exam questions with 75% accuracy. | Benchmarking Large Language Models for Quebec Insurance: From Closed-BarXiv:2603.07825 · 2026 |
| Assessment Design in the AI Era | Researchers from the Weizmann Institute applied a psychometric method for finding unfair exam questions — and turned it against six LLMs on data from 5700+ people. | Assessment Design in the AI Era: A Method for Identifying Items FunctiarXiv:2603.23682 · 2026 |
| Auditing the Reliability of Multimodal Generative Search | Researchers checked almost 12,000 statements from Gemini 2.5 Pro tied to YouTube videos: from 4% to 19% of statements are not confirmed by the video. | Auditing the Reliability of Multimodal Generative SearcharXiv:2604.00944 · 2026 |
| ClawSafety | 'Update the file'—is a trigger: the model searches for permission from 4–5 sources. | ClawSafety: "Safe" LLMs, Unsafe AgentsarXiv:2604.01438 · 2026 |
| Train Yourself as an LLM | Participants after one session had 42% less chance of falling for AI persuasion in all three scenarios at once. | Train Yourself as an LLM: Exploring Effects of AI Literacy on PersuasiarXiv:2604.02637 · 2026 |
| Vulnerability Detection with Interprocedural Context in… | Paradox: you add 'clarifying' materials—you lose 25 percentage points of accuracy. | Vulnerability Detection with Interprocedural Context in Multiple LanguarXiv:2604.08417 · 2026 |
| Do AI Coding Agents Log Like Humans? | In 67% of cases, the phrase 'add logging' in the prompt is a request into the void. | Do AI Coding Agents Log Like Humans? An Empirical StudyarXiv:2604.09409 · 2026 |
| RealVuln | Up to 99.5% of industrial security scanner alerts are false alarms. | RealVuln: Benchmarking Rule-Based, General-Purpose LLM, and Security-SarXiv:2604.13764 · 2026 |
| Diagnosing LLM Judge Reliability | Paradox: on average, an LLM evaluator gives contradictory ratings in less than 5% of cases. | Diagnosing LLM Judge Reliability: Conformal Prediction Sets and TransiarXiv:2604.15302 · 2026 |
| Human-Guided Harm Recovery for Computer Use Agents | Paradox: from 1150 real comparisons it turned out that people in crisis choose a quick targeted plan more often than a comprehensive one. | Human-Guided Harm Recovery for Computer Use AgentsarXiv:2604.18847 · 2026 |
| Serialisation Strategy Matters | The trick: data presentation format changes answer completeness much more than it seems – narrative text with explicit section headers gives small models around 7-8B parameters 19 percentage points higher completeness compared to raw JSON simply because the model stops spending effort parsing structure and immediately works with content. | Serialisation Strategy Matters: How FHIR Data Format Affects LLM MedicarXiv:2604.21076 · 2026 |
| Persona-Conditioned Adversarial Prompting | The method combines a specific persona (role with context) and a request submission strategy from 12 variants. | Persona-Conditioned Adversarial Prompting: Multi-Identity Red-Teaming arXiv:2605.11730 · 2026 |
| A-ProS | 2.9–3.5 times fewer repeat errors from a single solution: do not close the chat between iterations. | A-ProS: Towards Reliable Autonomous Programming Through Multi-Model FearXiv:2605.18073 · 2026 |
| An Empirical Study of Privacy Leakage Chains via Prompt… | Result: 72% successful attacks versus 3% with old methods — 24 times more effective. | An Empirical Study of Privacy Leakage Chains via Prompt Injection in BarXiv:2605.18133 · 2026 |
| Automated Grading of Handwritten Mathematics Using Vision-… | Actually in 87% of cases the problem is bad photo, not evaluation logic. | Automated Grading of Handwritten Mathematics Using Vision-Capable LLMsarXiv:2605.19043 · 2026 |
| Beyond Ideal Instruction | Top LLMs handle less than 40% of real tasks. | Beyond Ideal Instruction: A Comprehensive Framework for Evaluating LLMarXiv:2606.03318 · 2026 |
| VCIFBench | 8 requirements in one prompt — and the probability of fulfilling them all drops to 7.5%. | VCIFBench: Evaluating Complex Instruction Following for Video UnderstaarXiv:2606.04588 · 2026 |
| Mutation Without Variation | In 87% of iterative chains, LLM changes words but freezes the text skeleton within a few steps — 93% of so-called 'improvements' are just rearrangements of the same blocks in different positions. | Mutation Without Variation: Convergence Dynamics in LLM-Driven ProgramarXiv:2606.05408 · 2026 |
| Evaluating Stochastic Collapse and Implicit Bias in… | Paradox: Claude Sonnet in tests chose one of four equivalent options in 97% of cases — against the expected 25%. | Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large LarXiv:2606.05874 · 2026 |
| How Seemingly Inconsequential Design Choices Dictate… | From 15% to 44% accuracy — without changing the model, without retraining, only by changing the request format. | How Seemingly Inconsequential Design Choices Dictate Performance of LLarXiv:2606.12407 · 2026 |
| From Shield to Target | Schema Lock allows you to forcibly switch the model into a mode of methodical execution: get a complete breakdown by your categories, without omissions and vague "on the other hand." Add an analytical scheme with categories and evaluation criteria to the prompt — the model's attention is fixed on the headings 9.6 times stronger, the uncertainty of each token falls by half. | From Shield to Target: Denial-of-Service Attacks on LLM-Based Agent GuarXiv:2606.14517 · 2026 |
| Teaching agentic AI to learn expert reasoning for rare… | The same model, the same questions — but accuracy on ultra-rare diseases jumps from 10.7% to 60.7%. | Teaching agentic AI to learn expert reasoning for rare disease diagnosarXiv:2606.16149 · 2026 |
| OTTER | The OTTER method allows you to find the specific words that trigger alarms in the moderation system and replace them with neutral synonyms without losing meaning by measuring each word's contribution to the toxicity score, finding the top-5 hot words and replacing them so the request's meaning does not change and the surface signal disappears, resulting in success rates from 7% to 84%. | OTTER: A Red-Teaming System for Toxicity-Evading Jailbreak Prompt OptiarXiv:2606.21077 · 2026 |
| The Language-Energy Divide | Cyrillic creates 2.5 times more tokens than Latin script; Kazakh or Tibetan inflates the request up to 179 times. | The Language-Energy Divide: Measuring Energy Costs of Multilingual LLMarXiv:2606.21869 · 2026 |
| Do vision-language models search like humans? | Paradox: people take 2.8 times longer to become convinced that 'there is nothing' — than to find the target. | Do vision-language models search like humans? Reasoning tokens as a rearXiv:2606.25066 · 2026 |
| Agentic AI-Powered Re-Identification | 94.7% accuracy where deanonymization is possible in principle. | Agentic AI-Powered Re-Identification: An Emerging, Scalable Threat to arXiv:2606.27936 · 2026 |
| An Empirical Analysis of Factual Errors in Human-Written… | This is precisely why the best model in the study finds only half of real errors — 52% F1. | An Empirical Analysis of Factual Errors in Human-Written Text and its arXiv:2606.27959 · 2026 |
| SpreadsheetBench 2 | The best models—Claude, GPT, Gemini—handle complex business tables correctly only in 35% of cases. | SpreadsheetBench 2: Evaluating Agents on End-to-End Business SpreadshearXiv:2606.29955 · 2026 |
| Attractor States Emerge in Multi-Turn LLM Conversations | The difference is noticeable after 15-20 messages: without an anchor the model becomes itself, with an anchor it stays your character longer. | Attractor States Emerge in Multi-Turn LLM ConversationsarXiv:2606.30571 · 2026 |
| When Can You Debias an LLM Judge? | GPT-4o-mini chose the longer text in 100% of cases — with absolutely identical content of both variants. | When Can You Debias an LLM Judge? Identifiability Limits, a Test, and arXiv:2607.02104 · 2026 |
| Compression, structure, and executor capability | +27 percentage points came from only one thing — changing the executor model. | Compression, structure, and executor capability: a controlled real-cosarXiv:2607.03048 · 2026 |
| Human and LLM Collaboration for Accelerated Materials… | Score on known tasks — 75% for the model versus 83% for the expert. | Human and LLM Collaboration for Accelerated Materials Synthesis and DiarXiv:2607.07604 · 2026 |
| Which Values Do LLMs Confuse? | The ranked-response method makes it possible to find the correct answer even when the model got it wrong on the first attempt — in 68% of cases it is hidden in the 2-3rd place on the list. | Which Values Do LLMs Confuse? A Schwartz-Based Recognition StudyarXiv:2607.20270 · 2026 |
| Estimating the Geopolitical Preferences of Large Language… | The spread is huge: GPT-5 agrees with 97% of resolutions, DeepSeek — only with 38%. | Estimating the Geopolitical Preferences of Large Language Models from arXiv:2607.25526 · 2026 |
| Shorter Reasoning, Earlier Answers? | Tell the model a token limit — and get zero accuracy gain, although reasoning honestly shortened by 12-17%. | Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning InterfaarXiv:2608.03401 · 2026 |
| Your Prompt Is Not the Only Prompt | The method makes it possible to find this hidden conflict and fix it before it quietly crashes accuracy — from 100% to 73% or from 52% to 7% without a single error in the logs. | Your Prompt Is Not the Only Prompt: How Much Do LLMs Weight StructuredarXiv:2608.08254 · 2026 |
| LLM within MCP Matters | Three techniques in the prompt text — a directive, examples of correct and incorrect behavior, a warning in the tool description — return accuracy above 86% where without them it drops to 15%. | LLM within MCP Matters: Measuring Inefficient Resource Utilization DriarXiv:2608.08467 · 2026 |
| Similarity Gates Approve Reversals | 96% match — this is the automatic verification score given to a pair of phrases "do not give the patient the drug" and "give the drug". | Similarity Gates Approve Reversals: A Validity Audit of Embedding-CosiarXiv:2608.10216 · 2026 |
| Can Frontier LLMs Match Natively Multimodal Embeddings? | The method makes it possible to find the right photo among visually similar candidates (up to 25 pieces) by text description — without installing a separate search tool, just with one prompt in Claude or GPT. | Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison oarXiv:2608.11343 · 2026 |
| Automatic or Controlled? | This explains the frequent pain of long dialogues: you discussed a project for 40 minutes, got distracted by 15 messages — and the model confuses details or loses them entirely. | Automatic or Controlled? Repetition Priming Reveals Divergent ProcessiarXiv:2608.14681 · 2026 |
| What Aggregate Scores Miss | Up to 8% of your tasks can start malfunctioning after a model update — and the report will beautifully write '+7% accuracy'. | What Aggregate Scores Miss: Measuring Item-Level Regressions in CommerarXiv:2608.17719 · 2026 |
| Retrieval-Augmented Generation (RAG) | Using LLM to generate code that structures data proved to be up to 90% faster and 85% cheaper than making LLM analyze all data directly. | FastRAG: Retrieval Augmented Generation for Semi-structured DataarXiv:2411.13773 · 2024 |
| NEXT-EVAL | Converting a complex web page structure into a simple format "unique element identifier -> its text" (Flat JSON) increased data extraction accuracy from 10% to 95% and reduced the number of errors (hallucinations) by 30 times. | NEXT-EVAL: Next Evaluation of Traditional and LLM Web Data Record ExtrarXiv:2505.17125 · 2025 |
| LLMs Judging LLMs | Instead of rating on a scale of 1 to 5, the model answers a checklist of yes/no type questions, which eliminates instability and hidden biases, ensuring reproducible results. | LLMs Judging LLMs: A Simplex PerspectivearXiv:2505.21972 · 2025 |
| Optimizing Sequential Multi-Step Tasks with Parallel LLM… | Running multiple parallel attempts to solve a task can either significantly accelerate answer delivery (up to 2.2x) or improve final accuracy and success rate of solutions. | Optimizing Sequential Multi-Step Tasks with Parallel LLM AgentsarXiv:2507.08944 · 2025 |
| ChainEdit | ChainEdit increases models' ability for logical generalization by more than 30%, ensuring internal consistency of knowledge after editing. | ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing througharXiv:2507.08427 · 2025 |
| AGENTiGraph | AGENTiGraph allows creating chatbots with accuracy up to 95% on narrowly specialized topics without hallucinations. | AGENTiGraph: A Multi-Agent Knowledge Graph Framework for Interactive, arXiv:2508.02999 · 2025 |
| Making Slow Thinking Faster | Discovery: LLMs generate up to 80% of redundant, predictable steps in Chain-of-Thought reasoning — the model "thinks aloud," but most of these thoughts carry no informational value. | Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via SteparXiv:2508.03346 · 2025 |
| A Stitch in Time Saves Nine | Increases accuracy by 8.2% while reducing token usage by 41.6%. | A Stitch in Time Saves Nine: Proactive Self-Refinement for Language MoarXiv:2508.12903 · 2025 |
| LeanRAG | During search, it pulls relevant hierarchy through a lowest common ancestor algorithm, reducing noise by 46% and improving answer accuracy. | LeanRAG: Knowledge-Graph-Based Generation with Semantic Aggregation anarXiv:2508.10391 · 2025 |
| Collaborative Intelligence | Discovery: Cybersecurity specialists use GPT-4 not for content generation, but in 40% of cases — to explain complex technical strings (PowerShell commands, code fragments, network logs). | Collaborative Intelligence: Topic Modelling of Large Language Model usarXiv:2508.18488 · 2025 |
| Bias Mitigation Agent | Reduces bias by 81% without loss of information relevance. | Bias Mitigation Agent: Optimizing Source Selection for Fair and BalancarXiv:2508.18724 · 2025 |
| Vis-CoT | Increases accuracy by 24 percentage points. | Vis-CoT: A Human-in-the-Loop Framework for Interactive Visualization aarXiv:2509.01412 · 2025 |
| YouthSafe | Discovered: Standard safety filters in GPT-4, Claude, and other LLMs detect only 12-18% of hidden psychological risks for adolescents — emotional manipulation, dependency formation, boundary violations bypass moderation. | YouthSafe: A Youth-Centric Safety Benchmark and Safeguard Model for LaarXiv:2509.08997 · 2025 |
| InfoGain-RAG | Discovered: LLMs are catastrophically sensitive to information noise in context — even documents relevant by keywords can reduce answer accuracy by 40% if they contain contradictions or 'filler'. | InfoGain-RAG: Boosting Retrieval-Augmented Generation via Document InfarXiv:2509.12765 · 2025 |
| SKILL-RAG | The method allows you to get accurate answers from documents, even if they contain 80% garbage information. | SKILL-RAG: Self-Knowledge Induced Learning and Filtering for RetrievalarXiv:2509.20377 · 2025 |
| CIFLEX | This reduces computational costs 3-5 times, eliminates model confusion, and accelerates multi-stage request processing. | CIFLEX: Contextual Instruction Flow for Sub-task Execution in Multi-TuarXiv:2510.01239 · 2025 |
| TutorBench | The TutorBench benchmark revealed a critical problem: even top LLMs (GPT-4, Claude, Gemini) perform the tutor role only 56% of the time — models by default strive to provide a ready-made answer, which kills learning. | TutorBench: A Benchmark To Assess Tutoring Capabilities Of Large LanguarXiv:2510.02663 · 2025 |
| Training LLM Agents to Empower Humans | It divides tasks into automatable stages and human choice moments, increasing quality and AI suggestion acceptance by 31%. | Training LLM Agents to Empower HumansarXiv:2510.13709 · 2025 |
| David vs. Goliath | Autopilot developers faced a problem: even large LLMs poorly generate code from text descriptions (30-40% success). | David vs. Goliath: A comparative study of different-sized LLMs for codarXiv:2510.14115 · 2025 |
| StreamingThinker | The model breaks complex queries into logical blocks with intermediate conclusions, preventing loss of focus on initial instructions and reducing answer wait time by 60% while preserving reasoning quality. | StreamingThinker: Large Language Models Can Think While ReadingarXiv:2510.17238 · 2025 |
| LLMLogAnalyzer | Dividing the task into classification and execution improves answer accuracy by 30-40% on ROUGE-1 metrics by eliminating the lack-of-focus problem. | LLMLogAnalyzer: A Clustering-Based Log Analysis Chatbot using Large LaarXiv:2510.24031 · 2025 |
| DART | Achieves speedup up to 2.4x while maintaining or improving accuracy by 5-7%. | DART: Difficulty-Adaptive Reasoning Truncation for Efficient Large LanarXiv:2511.01170 · 2025 |
| Content & Reasoning Compression | RAG systems consume up to 60% of tokens on 'noisy' text and excessive agent reasoning — this is not a bug but the absence of information filtering mechanisms and efficiency control. | TeaRAG: A Token-Efficient Agentic Retrieval-Augmented Generation FramearXiv:2511.05385 · 2025 |
| Structured Clarification (SAGE-Agent Logic) | LLM agents fail 30-40% of tasks not due to model weakness, but because they make up missing details instead of clarifying ("call Alex" → agent calls a random Alex from contacts). | Structured Uncertainty guided Clarification for LLM AgentsarXiv:2511.08798 · 2025 |
| Graph-based Path Tracking | The system tracks logical chains between facts, cutting off irrelevant branches, which increases answer accuracy by 40% compared to standard RAG for multi-stage requests. | NeuroPath: Neurobiology-Inspired Path Tracking and Reflection for SemaarXiv:2511.14096 · 2025 |
| Retrieval-Augmented Generation (RAG) | Research showed that GPT-4 agents "reading" HTML code from e-commerce website pages fail tasks 2-3 times more often than agents receiving cleaned text (RAG approach). | MCP vs RAG vs NLWeb vs HTML: A Comparison of the Effectiveness and EffarXiv:2511.23281 · 2025 |
| Multi-Agent Delegation | The system synthesizes the final answer from the best ideas of top agents, ensuring a quality increase of 8.4% compared to a single model. | ART: Adaptive Response Tuning Framework -- A Multi-Agent Tournament-BaarXiv:2512.00617 · 2025 |
| AgentBay | Allows seamless transfer of control between agent and operator when obstacles arise, increasing task completion success by 48%. | AgentBay: A Hybrid Interaction Sandbox for Seamless Human-AI InterventarXiv:2512.04367 · 2025 |
| IaC Generation with LLMs | The problem: basic GPT-4 succeeds in only 27% of cases. | IaC Generation with LLMs: An Error Taxonomy and A Study on ConfiguratiarXiv:2512.14792 · 2025 |
| Adversarial versification in portuguese as a jailbreak… | Manual poems give approximately 62% success rate for bypassing protections, automatic ones approximately 43%, some models over 90%. | Adversarial versification in portuguese as a jailbreak operator in LLMarXiv:2512.15353 · 2025 |
| A Plan Reuse Mechanism for LLM-Driven Agent | If 30% of requests are similar, you can reuse old plans and reduce | A Plan Reuse Mechanism for LLM-Driven AgentarXiv:2512.21309 · 2025 |
| Agentic LLMs as Powerful Deanonymizers | A researcher took a public dataset of scientist interviews from Anthropic (1,250 interviews, of which 125 with scientists) and showed: an ordinary LLM agent with web search was able to deanonymize 25% of participants—based on descriptions of their research projects, the model found specific publications, authors, and sometimes uniquely identified the person. | Agentic LLMs as Powerful Deanonymizers: Re-identification of ParticipaarXiv:2601.05918 · 2026 |
| Retrieval-Augmented Generation (RAG) | Reduces vector dimensionality from 3072 to 110 measurements, providing 60-fold search acceleration and 28-fold index volume reduction with minimal accuracy loss. | PCA-RAG: Principal Component Analysis for Efficient Retrieval-AugmentearXiv:2504.08386 · 2025 |
| Proof-Carrying Numbers (PCN) | Discovery: LLMs distort numbers even with accurate data before their eyes — the model generates 'about 6%' instead of '5.7%' because it sounds more natural for a language model. | Proof-Carrying Numbers (PCN): A Protocol for Trustworthy Numeric AnswearXiv:2509.06902 · 2025 |
| In-Context Representation Hijacking | Doublespeak is a technique showing how LLMs redefine the internal meaning of a word through a series of contextual examples where if you systematically replace word A with word B across 5 to 10 sentences the model begins to internally interpret word B as word A even if on the surface they are unrelated, for example replacing "bomb" with "carrot" in several examples causes the request "how to build a carrot?" to be read by the model as "how to build a bomb?". | In-Context Representation HijackingarXiv:2512.03771 · 2025 |
| Trust Propagation Scoring | Integration of ClaimTrust into a RAG system increased the perceived quality and reliability of LLM answers by 11.2%. | ClaimTrust: Propagation Trust Scoring for RAG SystemsarXiv:2503.10702 · 2025 |
| SymRAG | It allows processing acceleration of 10-12 times without loss of accuracy. | SymRAG: Efficient Neuro-Symbolic Retrieval Through Adaptive Query RoutarXiv:2506.12981 · 2025 |
| Reasoning about Intent for Ambiguous Requests | on ambiguous SQL questions (vs 72.7% for SFT) and | Reasoning about Intent for Ambiguous RequestsarXiv:2511.10453 · 2025 |
| Tree-of-Code (ToC) | This increases accuracy by 20% and reduces iterations by 4 times. | Tree-of-Code: A Tree-Structured Exploring Framework for End-to-End CodarXiv:2412.15305 · 2024 |
| From Documents to Dialogue | Key result: This approach (KG-RAG) reduced the number of irrelevant answers by 52% and increased the number of completely relevant answers by 88% compared to the standard RAG approach | From Documents to Dialogue: Building KG-RAG Enhanced AI AssistantsarXiv:2502.15237 · 2025 |
| Self-Truncation Best-of-N (ST-BoN) | This approach reduces GPU memory usage by more than 90% and response time by 50% while achieving equal or even better answer quality compared to exhaustive enumeration of N variants. | Sampling-Efficient Test-Time Scaling: Self-Estimating the Best-of-N SaarXiv:2503.01422 · 2025 |
| Self-Consistency | Majority voting among 5 different settings (deterministic, balanced, and exploratory) is used to select the final answer, improving accuracy from 66% to 75%. | Leveraging LLMs for Predicting Unknown Diagnoses from Clinical NotesarXiv:2503.22092 · 2025 |
| Hidden State Forensics | Malicious requests create unique 'fingerprints' in neural activity, allowing detection of jailbreaks and hallucination-inducing requests with 95% accuracy before response generation. | Exposing the Ghost in the Transformer: Abnormal Detection for Large LaarXiv:2504.00446 · 2025 |
| Sentinel | It is trained on thousands of examples of malicious attacks and determines with 98.7% accuracy whether a user is trying to "hack" the system through a prompt. | Sentinel: SOTA model to protect against prompt injectionsarXiv:2506.05446 · 2025 |
| Context Filtering (ATF) | Pre-filtering data according to a specific question allows LLMs to provide more accurate answers, especially on complex and unfamiliar datasets, reducing the volume of information by up to 70%. | What to Keep and What to Drop: Adaptive Table Filtering FrameworkarXiv:2506.23463 · 2025 |
| Activation Steering for Chain-of-Thought Compression | This method, requiring no model retraining, allows reducing the length of reasoning chains (CoT) to 67%, which dramatically accelerates response generation. | Activation Steering for Chain-of-Thought CompressionarXiv:2507.04742 · 2025 |
| Smoothie-Qwen | Requires no retraining, maintains quality on target languages, reduces unwanted responses in the dominant language by 95%. | Smoothie-Qwen: Post-Hoc Smoothing to Reduce Language Bias in MultilingarXiv:2507.05686 · 2025 |
| ShoppingBench | Discovered: LLMs catastrophically fail on tasks with multiple simultaneous conditions — even GPT-4 successfully completes less than 50% of requests like "find a product + stay within budget + apply coupon + verify attributes". | ShoppingBench: A Real-World Intent-Grounded Shopping Benchmark for LLMarXiv:2508.04266 · 2025 |
| Disentangling Reasoning Logic to Resolve Explicit Knowledge… | Increases accuracy by 15-20% when working with contradictions. | Disentangling Reasoning Logic to Resolve Explicit Knowledge ConflictsarXiv:2508.01273 · 2025 |
| Real-Time Detection of Hallucinated Entities in Long-Form… | A trained classifier (probe) recognizes patterns in hidden layers of the model characteristic of generating fictitious entities (names, dates, quotes), achieving 95% accuracy without accessing external sources. | Real-Time Detection of Hallucinated Entities in Long-Form GenerationarXiv:2509.03531 · 2025 |
| Certainty-Guided Reasoning in Large Language Models | The system periodically evaluates the probability of the current answer being correct: at high confidence (97%+) it returns the result, at low confidence — it continues reasoning through a special Wait token. | Certainty-Guided Reasoning in Large Language Models: A Dynamic ThinkinarXiv:2509.07820 · 2025 |
| FRIT | The method fine-tunes the model to generate only those reasoning chains where each step actually affects the final answer, increasing accuracy from 30% to 95%. | FRIT: Using Causal Importance to Improve Chain-of-Thought FaithfulnessarXiv:2509.13334 · 2025 |
| Failure Makes the Agent Stronger | This breaks the cycle of repeated errors and shifts the agent from inertial repetition mode to conscious correction mode, improving accuracy from 30% to 95%. | Failure Makes the Agent Stronger: Enhancing Accuracy through StructurearXiv:2509.18847 · 2025 |
| CompLLM | Accelerates processing 4x, eliminates the problem of forgetting information from the middle of a document, improves answer quality on ultra-long contexts. | CompLLM: Compression for Long Context Q&AarXiv:2509.19228 · 2025 |
| Explore Briefly, Then Decide | Reduces response length to 71% without losing accuracy, eliminating the 'overthinking' problem — excessive model deliberation. | Explore Briefly, Then Decide: Mitigating LLM Overthinking via CumulatiarXiv:2510.02249 · 2025 |
| Think Just Enough | Allows assessing reliability based on generation style rather than content, and stopping generation when confidence threshold is reached, saving 50-66% of tokens at 92-95% accuracy. | Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LarXiv:2510.08146 · 2025 |
| Retrieval-in-the-Chain | An LLM that immediately searches for an answer misses 40-60% of relevant documents — it doesn't have time to 'understand' the nuances of the query. | Retrieval-in-the-Chain: Bootstrapping Large Language Models for GeneraarXiv:2510.13095 · 2025 |
| Think Twice | LLM loses 40-60% assessment quality when trying to check text against 10+ criteria simultaneously — the model's attention scatters and it misses critical errors. | Think Twice: Branch-and-Rethink Reasoning Reward ModelarXiv:2510.23596 · 2025 |
| Semantically-Aware LLM Agent to Enhance Privacy in… | Preserves the quality of model responses while ensuring privacy, reducing errors 5 times compared to traditional anonymization methods. | Semantically-Aware LLM Agent to Enhance Privacy in Conversational AI SarXiv:2510.27016 · 2025 |
| Efficient Test-Time Retrieval Augmented Generation | With ET2RAG you can increase RAG system accuracy by 6-16% without model retraining — the method generates several short answers based on different combinations of retrieved documents and selects the best one through similarity voting | Efficient Test-Time Retrieval Augmented GenerationarXiv:2511.01059 · 2025 |
| Retrieval-Augmented Generation (RAG) | Using the ROBOTO2 system, you can reduce the time for assessing the quality of clinical studies from 30+ minutes to ~15-20 minutes, using LLM to automatically find relevant fragments in articles and provide preliminary answers to standardized questions about risks of systematic errors | ROBoto2: An Interactive System and Dataset for LLM-assisted Clinical TarXiv:2511.03048 · 2025 |
| Efficient Reasoning via Thought-Training and Thought-Free… | Ensures accuracy retention while reducing generation time and answer length by 3-5 times compared to traditional Chain-of-Thought. | Efficient Reasoning via Thought-Training and Thought-Free InferencearXiv:2511.03408 · 2025 |
| E-CARE | Increases search accuracy by 12-15%. | E-CARE: An Efficient LLM-based Commonsense-Augmented Framework for E-CarXiv:2511.04087 · 2025 |
| Learnable Prefix / Semantic Initialization | A small learnable prefix (0.003% of model size) can radically change LLM behavior towards safety without full retraining. | Patching LLM Like Software: A Lightweight Method for Improving Safety arXiv:2511.08484 · 2025 |
| TokenSqueeze | Allows reducing token count by 50% while maintaining answer quality. | TokenSqueeze: Performance-Preserving Compression for Reasoning LLMsarXiv:2511.13223 · 2025 |
| Is Vibe Coding Safe? | Main finding: Coding-agents (SWE-Agent, Claude Code) can implement new features in large projects — best result 61% functionally correct solutions. | Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated CodarXiv:2512.03262 · 2025 |
| SoK | Researchers created a unified framework for causality analysis at four levels: tokens (which words are critical), neurons (which 1-2% of neurons are responsible for safety), layers (in which layers is protection concentrated), representations (how embeddings encode safety boundaries). | SoK: a Comprehensive Causality Analysis Framework for Large Language MarXiv:2512.04841 · 2025 |
| Matching Ranks Over Probability Yields Truly Deep Safety… | At each generation step you look at the top-20 tokens, ignore the 'Sorry' token with 97% probability, select the 'instruction' token with probable... | Matching Ranks Over Probability Yields Truly Deep Safety AlignmentarXiv:2512.05518 · 2025 |
| MCP-SafetyBench | Result: all models are vulnerable, with attack success rates ranging from 29.8% to 48.2%. | MCP-SafetyBench: A Benchmark for Safety Evaluation of Large Language MarXiv:2512.15163 · 2025 |
| How to Train Your Advisor | The advisor analyzes the request and creates specific guidance on style, tone and format, increasing answer relevance by 40-60% in personalization tasks. | How to Train Your Advisor: Steering Black-Box LLMs with Advisor ModelsarXiv:2510.02453 · 2025 |
| Decision Routing & Grounded Code Generation | By combining speech recognition, LLM routing, and secure code execution, you can create a voice assistant for data analysis that understands questions in natural language and responds with charts, tables, or voice explanations — with 95.8% accuracy and ~1.5 second latency | A Multimodal Conversational Agent for Tabular Data AnalysisarXiv:2511.18405 · 2025 |
| AI-Enabled grading with near-domain data for scaling… | Researchers trained a BERT model sequentially on three related questions about DNA replication, transcription, and translation and discovered that a model which saw evaluation examples from the first question requires 2-3 times fewer examples for quality evaluation of the second and third questions. | AI-Enabled grading with near-domain data for scaling feedback with humarXiv:2512.04113 · 2025 |
| DRS | Increases reformulation accuracy from 23% to 70% for GPT-3.5. | DRS: Deep Question Reformulation With Structured OutputarXiv:2411.17993 · 2024 |
| Cyri | The authors prove that through such advanced prompt engineering, even a relatively small model achieves high accuracy (95%) in a complex task. | Cyri: A Conversational AI-based Assistant for Supporting the Human UsearXiv:2502.05951 · 2025 |
| ReAct (Reason + Act) | Outperforms standard RAG by 41% through active rather than passive information retrieval. | Search-R1: Training LLMs to Reason and Leverage Search Engines with RearXiv:2503.09516 · 2025 |
| Draft-and-Verify Prompting | Allows acceleration of generation by 2-3 times, sometimes improving quality due to the guiding effect of drafts. | Consultant Decoding: Yet Another Synergistic MechanismarXiv:2506.02391 · 2025 |
| RAG-R1 | Key result: The parallel query approach (multi-query parallelism) not only speeds up the answer process, but also significantly increases its accuracy (up to 13 | RAG-R1: Incentivizing the Search and Reasoning Capabilities of LLMs tharXiv:2507.02962 · 2025 |
| Disambiguation-Centric Finetuning Makes Enterprise Tool-… | This prevents impulsive incorrect decisions when choosing between similar tools or action options, reducing errors by 27-49% compared to the standard approach. | Disambiguation-Centric Finetuning Makes Enterprise Tool-Calling LLMs MarXiv:2507.03336 · 2025 |
| Odysseus | Odysseus is an attack on multimodal LLM systems like GPT-4o, Gemini, and Grok that hides a malicious request in an innocent image through steganography, a technique of embedding hidden information, so the system sees no explicit text and filters pass it through, the model generates an answer that is also hidden in the image, and the attacker decodes it locally, with success rates up to 99%. | Odysseus: Jailbreaking Commercial Multimodal LLM-integrated Systems viarXiv:2512.20168 · 2025 |
| The ORCA Benchmark | Top LLMs (ChatGPT-5, Claude, Gemini) correctly solve only 45-63% of everyday computational tasks — finance, physics, health. | The ORCA Benchmark: Evaluating Real-World Calculation Accuracy in LargarXiv:2511.02589 · 2025 |
| From User Interface to Agent Interface | LLM-agents for UI automation (application testing, AI assistants) spend 80-99% of all tokens on UI representations, DOM trees and accessibility trees describing interface structure. | From User Interface to Agent Interface: Efficiency Optimization of UI arXiv:2512.13438 · 2025 |
| Do You Really Need a GPU to Guard Your LLM? | Researchers showed that a simple classifier (text normalization + TF-IDF + Linear SVM) is more accurate and 10 times faster than advanced ShieldGemma. | Do You Really Need a GPU to Guard Your LLM? CPU-Class Classifiers and arXiv:2512.19011 · 2025 |
| HAFix | Adding the task evolution history to the prompt increases LLM solution accuracy by 45%, as the model receives context about failed attempts and code development direction. | HAFix: History-Augmented Large Language Models for Bug FixingarXiv:2501.09135 · 2025 |
| ALAS | Increases the accuracy of answers to current questions from 15% to 90% without human involvement. | ALAS: Autonomous Learning Agent for Self-Updating Language ModelsarXiv:2508.15805 · 2025 |
| Leveraging Vision Capabilities of Multimodal LLMs for… | Data extraction accuracy exceeding 90% has been achieved with positioning error of approximately 5% on synthetic and real scientific plots. | Leveraging Vision Capabilities of Multimodal LLMs for Automated Data EarXiv:2503.12326 · 2025 |
Reading this honestly
An effect measured on one benchmark with one model in one year is evidence, not a guarantee. Several widely repeated prompting results have shrunk or reversed on 2025-2026 reasoning models, which is why the technique pages state what the effect was measured on rather than presenting it as a property of the technique.
The table is also not a ranking. A large reported effect on a narrow benchmark is not better than a modest one on a broad task, and nothing here tells you which applies to what you are doing. That question is the reason the product exists.
Which of these applies to your task?Describe what you are doing and get a prompt built on the technique that fits. Free, no account.
Open the generator