Prompting technique

Rebuttal Prompting

SycEval: Evaluating LLM Sycophancy, arXiv:2502.08177 (2025) · read the paper

What it does

Rebuttal Prompting — a technique for correcting LLM errors through direct indication of inaccuracy. A simple rebuke ('you are wrong') is more effective than complex refutations with citations. Minimizes model sycophancy and increases answer accuracy, preventing regressive compliance where the model changes a correct answer to an incorrect one under pressure from authoritative-sounding arguments.

Task typemulti step
Shapemulti turn
Effortsingle step
Addressesinvented facts and citations; answers that are close but wrong; the same prompt giving different answers
Published2025-02-12
AuthorsAaron Fanous, Jacob Goldberg, Ank A. Agarwal et al.

When it helps

Use this technique when working with LLMs in dialogue mode when the model gives incorrect answers and correction is needed. Particularly effective in customer support scenarios, educational applications, chatbots, and fact-checking systems. Apply when developing prompts to minimize model sycophancy and increase answer accuracy. Critical for systems requiring high information reliability and the model's ability to acknowledge errors. Suitable for testing and evaluating resist

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 evaluating and correcting responses based on the SycEval technique, specifically focusing on "Rebuttal Prompting" to improve accuracy and reduce sycophancy.

Your goal is to demonstrate how to effectively correct an LLM's incorrect response by using a direct, simple rebuttal, rather than complex or fabricated arguments.

**Scenario:**
An LLM has provided an incorrect answer to a user's request. You need to construct a prompt that guides the user on how to correct the LLM.

**Instructions:**
1.  **Identify the core mistake:** Assume the LLM has made a factual error or provided an irrelevant response.
2.  **Formulate a "Simple Rebuttal":** Create a direct, concise statement indicating the LLM is wrong.
3.  **Provide clear correction criteria:** Specify what the LLM needs to do to generate a correct response.
4.  **Structure the prompt:** Use clear sections for Role, Context, and Task.
5.  **Use placeholders for user adaptation:** Mark sections where the user needs to insert their specific task and the LLM's incorrect response.

**Prompt Template:**

### Role
You are an AI assistant that helps users correct LLM responses using the SycEval "Rebuttal Prompting" technique. Your purpose is to guide users on how to provide direct feedback to improve LLM accuracy.

### Context
The user has received an incorrect or unsatisfactory response from an LLM. The goal is to correct this response efficiently by employing a "Simple Rebuttal" as recommended by SycEval. This technique avoids complex, potentially misleading arguments and focuses on direct feedback to prompt the LLM to re-evaluate its output.

### Task
1.  **Review the LLM's incorrect response:**
    ```
    [INSERT THE LLM'S INCORRECT RESPONSE HERE]
    ```

2.  **State the user's original request:**
    ```
    [INSERT THE USER'S ORIGINAL PROMPT/REQUEST HERE]
    ```

3.  **Construct the corrective prompt for the LLM:**
    Based on the SycEval technique, use a "Simple Rebuttal" to indicate the LLM's error. Follow this with clear, actionable instructions for correction.

    **Your corrective prompt should look like this:**

    "Regarding your previous response to the prompt: '[INSERT THE USER'S ORIGINAL PROMPT/REQUEST HERE]'

    **Your response is incorrect/unsatisfactory.**

    [Explain briefly *why* it's incorrect or unsatisfactory, e.g., "It contains factual inaccuracies about X," or "It did not address the core of the request regarding Y."]

    Please regenerate your response, adhering to the following corrected criteria:
    - [Specify the first correction criterion, e.g., "Ensure all factual claims about X are accurate and verifiable."]
    - [Specify the second correction criterion, e.g., "Focus on the user's specific requirement for Y."]
    - [Add any other necessary criteria for a correct response.]

    **Do not** attempt to justify your previous error or provide lengthy explanations. Simply provide a new, corrected response."

4.  **Explain the rationale:**
    Briefly explain why this "Simple Rebuttal" approach is effective according to SycEval, emphasizing that direct correction is better than complex arguments for improving LLM accuracy and avoiding sycophancy.

If this one does not fit, the two closest alternatives in the corpus are Debugging Prompts Framework and BloomWise 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 specialized in evaluating and correcting LLM outputs based on the SycEval methodology. Your task is to review a draft for unsupported claims.

**Context:**
The user has provided a draft containing various claims. The goal is to identify and flag any claims that lack sufficient evidence or cannot be supported.

**Technique: SycEval (Rebuttal Prompting)**
The core principle is that direct, simple rebuttals are more effective than complex, fabricated arguments when correcting an LLM.

**Task:**
Review the following draft. Identify claims that are not supported by evidence or are potentially inaccurate. For each unsupported claim, provide a direct rebuttal.

**Draft for Review:**
[User to insert draft content here]

**Instructions:**
1.  Read through the provided draft carefully.
2.  Identify any statement that appears to be an unsupported assertion, a generalization without evidence, or a potentially false claim.
3.  For each identified claim, formulate a **"Simple Rebuttal"**. This should be a direct, concise statement indicating the claim is problematic. Avoid providing lengthy explanations or fabricated evidence.
    *   Examples of Simple Rebuttals:
        *   "This claim is unsupported."
        *   "This statement is inaccurate."
        *   "This assertion lacks evidence."
        *   "This is not verifiable."
4.  Present your findings as a list of claims followed by their corresponding simple rebuttals.

**Output Format:**
List each unsupported claim and its rebuttal clearly.

**Example of expected output for a hypothetical claim:**
*   **Claim:** "All birds can fly."
    **Rebuttal:** "This claim is unsupported."

*   **Claim:** "The Earth is flat."
    **Rebuttal:** "This statement is inaccurate."

Begin the review process now.

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

Debugging Prompts FrameworkA systematic framework for diagnosing and preventing typical errors of language models in instruction understa…BloomWise PromptingPrompt structuring method based on Bloom's taxonomy, which guides the LLM through six levels of cognitive thin…RESS (Decomposition-then-Synthesis)A method of decomposing a complex task into a sequence of simple analytical steps followed by synthesis. Inste…Generative Self-Aggregation (GSA)A method for generating improved answers through self-aggregation: the model creates multiple different answer…

All techniques · Failure modes and fixes