Prompting technique

Stated Answer vs. Revealed Belief Prompting

Implicit Probabilistic Reasoning Does Not Reflect Explicit Answers in Large Language Models, arXiv:2406.14986 (2024) · read the paper

What it does

Technique comparing two LLM operation modes: analytical direct answer (Stated Answer) and free text generation (Revealed Belief). Reveals the gap between what the model 'knows' and how it actually generates text, uncovering hidden biases and irrelevant context influence.

Task typereasoning
Shapemulti turn
Efforttwo or three steps
Addressesthe same prompt giving different answers; invented facts and citations; answers that are close but wrong
Published2024-06-21
AuthorsManuel Mondal, Ljiljana Dolamic, Gérôme Bovet et al.

When it helps

Apply this technique when verifying LLM response reliability and identifying model biases. Particularly useful for critical decision-making scenarios requiring confirmation that the model truly understands the task, not just generating plausible text. Use for prompt testing, detecting model cognitive distortions, and improving production response quality.

A prompt you can paste

Generic by design: it applies the technique without knowing your task. Adapt the marked parts.

Prompt
You are an AI assistant tasked with demonstrating the "Stated Answer vs. Revealed Belief" prompting technique. Your goal is to create a reusable prompt template that a user can adapt for their specific needs.

**Technique:** Stated Answer vs. Revealed Belief Prompting

**Core Principle:** LLMs often "know" the correct answer or possess the necessary knowledge but fail to apply it in free-form text generation due to internal biases or irrelevant context. This technique forces the LLM to first engage in analytical "Stated Answer" mode before generating free-form "Revealed Belief" text.

**Prompt Structure:**

1.  **Role Definition:** Clearly define the AI's persona and expertise.
2.  **Context Setting:** Provide all necessary background information and constraints for the task.
3.  **Decomposition into Analytical Steps ("Stated Answer" Mode):** Break down the main task into a series of smaller, specific, analytical questions or commands. Each step should require the LLM to process information, make logical deductions, or extract specific data.
    *   Use clear numbering or bullet points for these steps.
    *   Specify the desired output format for each step (e.g., list, table, summary).
4.  **Synthesis Command ("Revealed Belief" Mode):** Instruct the LLM to generate the final output, explicitly stating that it should be based *only* on the information and analysis derived from the preceding steps. This prevents the LLM from falling back on its "instincts" and incorporating irrelevant patterns.
5.  **Final Output Formatting:** Specify the desired format for the overall response.

---

### **Reusable Prompt Template**

**You are a [Define AI Role/Persona, e.g., expert financial analyst, creative content strategist, meticulous researcher].**

**Your task is to [Clearly state the overall objective, e.g., analyze a business situation, generate creative marketing copy, summarize complex research].**

**Context:**
*   **[Provide detailed background information relevant to the task.]**
*   **[List any constraints, requirements, or specific criteria that must be met.]**
*   **[Include any relevant data or input that the AI needs to process, e.g., a news article, a user scenario, a product description.]**

**Execute the task using the following structured approach:**

**Step 1: [Specific Analytical Task 1]**
*   **Objective:** [Clearly state what the AI needs to do in this step, e.g., "Identify the key stakeholders involved."]
*   **Input:** [Specify the data or context from above to be used for this step.]
*   **Output Format:** [Define the expected output, e.g., "A bulleted list of stakeholders."]

**Step 2: [Specific Analytical Task 2]**
*   **Objective:** [e.g., "Analyze the potential risks associated with each stakeholder."]
*   **Input:** [Reference the output from Step 1 and/or relevant context.]
*   **Output Format:** [e.g., "A table with columns: 'Stakeholder', 'Potential Risk', 'Severity (Low/Medium/High)'.]

**Step 3: [Specific Analytical Task 3, continue as needed]**
*   **Objective:** [e.g., "Determine the primary motivations for each stakeholder."]
*   **Input:** [Reference previous steps/context.]
*   **Output Format:** [e.g., "A concise summary for each stakeholder."]

**(Add more analytical steps as required to fully break down the task)**

**Final Synthesis:**
Based *solely* on the analysis and information derived from Steps 1, 2, and 3 (and any subsequent steps), **[State the final output command, e.g., "generate a comprehensive strategy document", "write a compelling social media post", "create a detailed report"].**

**Final Output Requirements:**
*   [Specify overall formatting, tone, length, or any other final requirements.]
*   Ensure the final output directly addresses the main task objective.

If this one does not fit, the two closest alternatives in the corpus are Anti-Bias Prompting and Tabular Prompting, which target the same failure from a different angle.

Worked example

The same technique applied to a concrete job: check a draft for claims that cannot be supported. Use it as the pattern for your own case rather than as a finished artefact.

Worked example
You are an AI assistant designed to verify claims and ensure factual accuracy. Your task is to review a draft document and identify any claims that lack sufficient evidence or support.

**Context:**
The user has provided a draft document that needs to be checked for unsupported claims.

**Core Principle (Stated Answer vs. Revealed Belief):**
LLMs can "know" the correct answer but fail to use that knowledge in free-form generation, succumbing to biases or irrelevant context. To ensure accuracy, we must force the model to first engage in analytical reasoning ("Stated Answer") before generating the final output ("Revealed Belief").

**Task Breakdown:**

**Step 1: Identify Potential Claims (Analytical Phase - "Stated Answer")**
Read through the provided draft text. Identify and list all statements that present factual assertions, statistics, or specific claims that would require external validation or evidence. For each identified claim, briefly note its nature (e.g., statistical, historical, scientific, anecdotal).

**Step 2: Analyze Support for Each Claim (Analytical Phase - "Stated Answer")**
For each claim identified in Step 1:
*   **Assess Evidence:** Determine if the draft provides any supporting evidence, data, citations, or logical reasoning for the claim.
*   **Identify Gaps:** If evidence is missing or insufficient, explicitly state the nature of the gap (e.g., "No source cited," "Statistic lacks context," "Anecdotal evidence presented as fact").
*   **Formulate Verification Questions:** For claims lacking support, formulate specific questions that would need to be answered to verify the claim (e.g., "What is the source of this statistic?", "Can this historical event be corroborated?").

**Step 3: Synthesize Findings and Report (Generative Phase - "Revealed Belief")**
Based on the analysis in Steps 1 and 2, compile a report detailing the findings. The report should:
*   Clearly list all claims identified as potentially unsupported.
*   For each unsupported claim, explain *why* it is considered unsupported, referencing the evidence gaps identified in Step 2.
*   Provide the verification questions formulated in Step 2 for each unsupported claim.
*   Conclude with a summary of the overall confidence in the draft's factual accuracy based on the identified gaps.

**Output Format:**
Present your findings in a structured report using Markdown. Use clear headings for each section (e.g., "Identified Claims," "Unsupported Claims Analysis," "Verification Questions"). For the analysis of unsupported claims, use a list or table format.

Get this written for your actual task

Paste what you are trying to do and the corpus will be matched against it directly. Free, no account, about ten seconds.

Free · no signup · ~10s
0.00match confidence
single retrieval pass
Prompt for your task

      

That number is low on purpose, and it is real. It is the raw similarity of one retrieval pass: no specialist read the paper, no judge compared anything, the first plausible match won.

6,235techniques in the corpus
one of which is this page

Picking the right one for a specific task is the work, and it is the work GetDecision does.

This pageone technique, generic prompt
What you just ranone technique matched to your wording, nothing verified
Full runten specialists read the papers in full, a judge ranks the top three for your task and shows its reasoning, generation on the model you pick, saved to your history

See the top three for your taskTen specialists, a judge, and the reasoning shown. Free account, first run included.

Run the full analysis

Related techniques

Anti-Bias PromptingThe technique of neutral prompt formulation that excludes cognitive biases and user prejudice. Essence: avoid …Tabular PromptingA method of presenting data as a markdown table before submitting a request to an LLM. The tabular structure h…Multilingual Ensemble PromptingThe method involves translating a complex query into 3-4 different languages, obtaining separate answers from …Debate-to-Detect (D2D) / Multi-Agent Structured DebateThe method transforms an analysis task into structured debates between multiple AI agents with opposite roles.…

All techniques · Failure modes and fixes