💬 NLP

Easier to Mislead Than to Correct: Harmful and Beneficial Revision in LLM Conformity

This study reveals that large language models in multi-agent systems are significantly more susceptible to being misled by peer consensus and authority labels than they are to being corrected, and that standard reasoning interventions fail to reliably mitigate this harmful conformity.

Original authors: Jiaming Qu, Lucheng fu, Yibo Hu

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaming Qu, Lucheng fu, Yibo Hu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Peer Pressure" Problem

Imagine you are taking a difficult trivia quiz. You are pretty sure the answer to a question is B. But then, six other people in the room all shout, "No, it's definitely A!"

This paper studies what happens when Artificial Intelligence (AI) models find themselves in that exact situation. The researchers wanted to know: Is it easier to trick a smart AI into changing a right answer to a wrong one, or to help a confused AI change a wrong answer to a right one?

The short answer is: It is much easier to trick the AI.

The Experiment: The AI in a Classroom

The researchers set up a "classroom" scenario for four different AI models (like digital students).

Round 1: The AI answers a question on its own.
The Twist: The AI is shown the answers of six "classmates" (simulated peers).
Round 2: The AI gets to change its answer if it wants to.

The researchers played with two main variables to see how they influenced the AI:

The Crowd: Did all the classmates agree on the same answer? (Sometimes they were all right, sometimes all wrong, sometimes mixed).
The Title: Did some classmates have fancy titles like "Team Leader" or "Research Director"?

The Three Big Discoveries

1. The "Bad News" Travels Faster Than the "Good News"

This is the paper's most important finding.

The Scenario: If the AI got the answer right initially, but all six classmates said it was wrong, the AI changed its answer to the wrong one 63% of the time.
The Comparison: If the AI got the answer wrong initially, but all six classmates said it was right, the AI only fixed its mistake 52% of the time.

The Analogy: Imagine you are holding a heavy, correct rock. If six people push you, it's very easy to knock the rock out of your hand and make you drop it (misleading). But if you are holding a broken rock and six people try to hand you a new, correct one, it's much harder for you to let go of the broken one and grab the new one (correcting).

The Takeaway: It is much easier to confuse a smart AI with a group of wrong answers than it is to fix a confused AI with a group of right answers.

2. The "Boss" Effect

The researchers also gave some classmates titles like "Team Leader."

The Result: When the AI saw a "Team Leader" pick an answer, it was more likely to change its mind to match that leader.
The Catch: It didn't matter if the leader was right or wrong. If the "Team Leader" said the answer was "A" (even if "A" was wrong), the AI was more likely to say "A."

The Analogy: It's like a student in a classroom who ignores the facts and just copies the answer of the teacher's favorite student, even if that student is guessing. The AI trusts the title more than the truth.

3. "Thinking Harder" Doesn't Always Help

The researchers tried to fix this problem by giving the AI two special instructions:

Chain-of-Thought: "Think step-by-step before answering."
Reflect-and-Revise: "Look at your answer again and think about whether you should change it."

The Result: These tricks didn't work the way we hoped.

Thinking Step-by-Step: This actually made the AI less likely to fix its mistakes. If the AI was wrong and the group was right, the AI would stick to its own wrong reasoning instead of listening to the group.
Reflecting: This made the AI very stubborn. It would just refuse to change its answer at all, whether that meant it kept a wrong answer or kept a right one.

The Analogy: It's like telling a stubborn person, "Think about it!" They might think about it, but instead of realizing they were wrong, they just convince themselves even more that they were right.

What Does This Mean for the Future?

The paper concludes that if we build systems where many AIs talk to each other (like a team of robots solving a problem), we cannot just let them vote on the answer.

Don't just count votes: If 5 AIs say "A" and 1 says "B," the group shouldn't automatically pick "A." The "A" answer might be a group hallucination (a shared mistake).
Don't trust titles: Just because one AI is labeled "Expert" doesn't mean it's right.
Check the work: Instead of just agreeing with each other, the AIs need to check the facts against the original question, rather than just listening to each other.

In short: AI is very good at following the crowd, even when the crowd is wrong, and it is surprisingly bad at fixing its own mistakes when the crowd is right.

Technical Summary: Easier to Mislead Than to Correct

Problem Statement

As Large Language Models (LLMs) are increasingly deployed in multi-agent systems (MAS), they frequently encounter scenarios where they must review and respond to the outputs of other agents. A critical risk in these settings is conformity: an LLM may abandon its own correct answer simply because a majority of peers agree on a different one. While prior research has established that LLMs tend to revise toward majority answers, it remains unclear whether peer influence is equally effective at correcting initially wrong models as it is at misleading initially correct ones. Furthermore, it is unknown whether generic reasoning interventions (e.g., Chain-of-Thought) can reliably distinguish between beneficial and harmful revisions.

Methodology

The authors conducted a controlled study across four open-weight LLMs (Qwen2.5-7B, Mistral-7B, Gemma-2-9B, Llama-3.1-8B) and seven QA/reasoning datasets (including BBH, MMLU-Pro, ARC-Challenge, and TruthfulQA). The experimental protocol involved a two-round process:

Round 1: The model answers a multiple-choice question independently and reports a self-confidence score (1–10).
Round 2: The model is presented with a prompt containing six simulated peer responses and asked to reconsider the question.

The study manipulated two social cues grounded in social-psychology theory:

Consensus Structure: Three conditions were tested: Mixed (3 peers correct, 3 wrong), All-Correct (6 peers correct), and All-Wrong (6 peers wrong).
Authority Labels: Peers were either unlabelled or assigned an organizational role (e.g., "team leader," "research director").

The study addressed three Research Questions (RQs):

RQ1 (Revision Quality): How do consensus structure and authority labels affect the rates of beneficial revision (wrong $\to$ correct) versus harmful revision (correct $\to$ wrong)?
RQ2 (Cue Strength): How do revision rates change as the number of committed peers or authority-labeled peers increases?
RQ3 (Reasoning Interventions): Do Chain-of-Thought (CoT) and "reflect-then-revise" prompts reduce harmful revision while preserving beneficial revision?

Key Contributions and Results

1. Asymmetry in Conformity (RQ1)

The central finding is that it is significantly easier to mislead an LLM than to correct it.

Harmful vs. Beneficial: When all peers agreed on a wrong answer, harmful revision rates surged from a baseline of 15.6% to 62.9% (+47.3 percentage points). In contrast, when all peers agreed on a correct answer, beneficial revision rates only increased from 32.7% to 51.5% (+18.8 percentage points).
Magnitude: Regression analysis indicated that the effect of all-wrong peers on harmful revision was five times larger than the effect of all-correct peers on beneficial revision (Odds Ratio: 28.5 vs. 5.2).
Authority: Authority labels increased the likelihood of choosing an endorsed answer, but their effect size was smaller than that of peer consensus. Crucially, authority labels influenced models regardless of whether the endorsed answer was correct or incorrect.

2. Strength of Social Cues (RQ2)

Committed Peers: Revision rates increased as more peers committed to a specific answer, but the effect was sublinear (diminishing returns), consistent with Social Impact Theory.
Authority Labels: As the number of authority-labeled peers increased, models were more likely to revise toward the authority-endorsed answer. This occurred whether the authority endorsed the correct or wrong answer, with a similar magnitude of effect in both scenarios.

3. Ineffectiveness of Generic Reasoning Interventions (RQ3)

Standard prompt-based interventions failed to selectively mitigate harm without sacrificing benefit:

Chain-of-Thought (CoT): While CoT reduced harmful revision under all-wrong peer conditions (from 63.9% to 36.5%), it simultaneously suppressed beneficial revision across all conditions. Models became less likely to accept correct peer corrections.
Reflect-Then-Revise: This intervention made models uniformly more conservative. It reduced harmful revision but also significantly reduced beneficial revision (e.g., dropping from 52.7% to 33.4% under all-correct peers). It acted more as a "partial reset" to the initial answer rather than a verifier of correctness.

Significance and Claims

The paper argues that the core risk for multi-agent LLM systems is not merely that models conform, but that wrong peer agreement propagates errors more effectively than correct peer agreement repairs them.

The authors claim that:

Aggregation is insufficient: Simply aggregating peer answers (e.g., via voting or majority rule) is dangerous because it amplifies errors more than it corrects them.
Authority is not a safeguard: Role labels (e.g., "expert") do not guarantee correctness and can be used to manipulate model outputs regardless of the truth.
Reasoning prompts are not a panacea: Generic interventions like CoT or reflection do not reliably separate harmful from beneficial influence; they often suppress the model's ability to learn from correct peers.

Consequently, the authors conclude that multi-agent systems should be designed to verify peer answers against task evidence rather than simply aggregating them. Future systems must move from asking "which agents agree?" to "which answer is best supported by evidence?" to prevent the cascading of errors in collaborative LLM environments.