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

Abstract: LLMs are increasingly used in multi-agent systems, where they see and respond to other agents' answers. A key risk is conformity: a model may abandon its own answer simply because others agree on a different one. Prior studies show that LLMs often revise toward a majority answer, but it remains unclear whether these revisions help correct mistakes as often as they introduce new errors. In this paper, we conduct a controlled study in which an LLM first answers a question, then sees simulated peer responses before making a final decision. We manipulate two social cues: consensus structure and authority labels assigned to peers, and measure how they influence beneficial and harmful revisions. Across four open-weight LLMs and seven QA datasets, we find that peer agreement makes it much easier to mislead initially correct models than to correct initially wrong ones. Authority labels make models more likely to choose the endorsed answer, regardless of whether it is correct. More concerningly, generic reasoning interventions such as chain-of-thought and reflection do not reliably reduce harmful revision while preserving beneficial revision. These findings suggest that multi-agent LLM systems should verify peer answers rather than simply aggregate them.

• The paper shows that unanimous wrong peers increased harmful revision from 15.6% to 62.9%, while unanimous correct peers increased beneficial revision from 32.7% to 51.5% across four models and seven QA datasets.
• Authority labels shifted model answers regardless of correctness, with committed peers producing graded but diminishing influence and authority cues increasing compliance by roughly 28–30 percentage points.
• The paper finds that chain-of-thought and reflect-then-revise reduce harmful conformity but also suppress beneficial correction, indicating that multi-agent systems need evidence-based verification rather than vote aggregation.

Overview

This paper presents a controlled experimental study of conformity in LLMs, asking whether peer influence is equally effective at correcting initially wrong answers and at misleading initially correct ones. The authors define beneficial revision (wrong → correct) and harmful revision (correct → wrong) and compare their rates under manipulated social cues. The central finding is stark: unanimous wrong peer agreement raises harmful revision far more sharply than unanimous correct agreement raises beneficial revision. Across four open-weight models and seven QA datasets, the paper further shows that authority labels shift model answers regardless of correctness, and that generic reasoning interventions—chain-of-thought (CoT) and reflect-then-revise—do not selectively suppress harmful revision while preserving beneficial revision (2606.01637).

The study is motivated by the growing deployment of multi-agent LLM systems (e.g., agent debate and LLM-as-a-Judge pipelines) in which agents see and respond to one another's outputs. Prior work has established that LLMs follow majorities, defer to confident peers, and respond to "expert" labels, but the relative strength of misleading versus corrective influence had not been directly compared within a single controlled design.

Experimental design

The protocol uses a two-round pipeline. In Round 1, the model answers a multiple-choice question independently and reports confidence on a 1–10 scale with explicit anchors. In Round 2, the model sees six simulated peer responses in its chat history and answers again. Peer utterances are drawn from fixed paraphrase templates; authority conditions tag peers with roles such as "team leader" or "research director." Peer names are sampled from a balanced pool of 40 names to reduce demographic confounds. All experiments use greedy decoding with three random seeds, and analyses employ logistic and linear mixed-effects regressions with a random intercept per question instance.

Three research questions structure the manipulations:

• RQ1 (revision quality) crosses consensus structure (mixed: 3 correct + 3 distinct wrong peers; all-correct; all-wrong) with authority-label presence.
• RQ2 (cue strength) varies committed-peer count $n_\text{com} \in \{0,2,4,6\}$ and authority-label count $n_\text{auth} \in \{1,\dots,5\}$.
• RQ3 (interventions) applies CoT prompting and reflect-then-revise to all RQ1 perturbations.

Evaluation covers Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, Gemma-2-9B-Instruct, and Llama-3.1-8B-Instruct on four BBH tasks plus MMLU-Pro, ARC-Challenge, and TruthfulQA, totaling 2,500 instances.

RQ1: Wrong agreement misleads more than correct agreement helps

The headline result is an asymmetry measured against a shared mixed-peer baseline. All-wrong peers raise harmful revision from 15.6% to 62.9% (+47.3 pp), whereas all-correct peers raise beneficial revision only from 32.7% to 51.5% (+18.8 pp). Regression analysis confirms this: the odds ratio for all-wrong peers on harmful revision is 28.5 versus 5.2 for all-correct peers on beneficial revision (both $p<.001$)—a roughly fivefold difference in effect size. This asymmetry holds across all four models and all seven tasks, though magnitudes vary considerably: Mistral-7B shows baseline harmful revision of 51.7% even under mixed peers, while Gemma-2-9B rarely revises at all (beneficial revision under all-correct peers is only ~5%).

Two explanations are offered, both acknowledged as partial. First, peer messages contain answer labels but no rationales, so wrong agreement may induce doubt while correct agreement supplies insufficient task evidence—an interpretation consistent with misinformation-correction literature showing that bare corrective answers do not resolve false beliefs. Second, the two measures are computed on different question subsets: beneficial revision is measured on harder items the model initially missed, which may cap how much peers can help. The authors state this confound plainly rather than claiming the asymmetry is purely psychological.

Authority labels have a smaller but significant effect: under mixed peers, adding one authority label raises harmful revision from 15.6% to 20.6% while slightly lowering beneficial revision. Notably, reported confidence does not rise when models conform—$\Delta C$ is more negative under all-agree conditions than under mixed peers for both revision types—indicating that peer pressure changes answers without increasing stated certainty.

A follow-up analysis adds a second failure mode: error convergence. Under all-wrong peers, 56.8% of initially wrong answers switch to the agreed wrong answer (versus 20.7% under mixed peers), meaning unanimous wrong peers pull already-wrong models onto a single shared error rather than enabling recovery.

RQ2: Social cues act in graded, sublinear fashion

RQ2a shows that revision increases with the number of committed peers—from 25.3% when no peer commits to 34.9% when all six commit—with significant positive linear and negative quadratic terms, consistent with diminishing marginal influence predicted by Social Impact Theory. The $n_\text{com}=0$ condition (all six peers express uncertainty) yields lower revision than any RQ1 condition, ruling out prompt-length as the driver of the RQ1 effects.

RQ2b shows that authority-aligned revision—the rate of switching to the option endorsed by authority-labeled peers—rises monotonically but sublinearly with label count, by approximately +30 pp when authorities endorse the correct answer and +28 pp when they endorse a wrong answer. The near-symmetry of these effects is the key point: authority labels increase compliance irrespective of correctness, making role tags a potential low-cost manipulation channel rather than a reliability signal.

RQ3: Reasoning interventions fail selectively

Neither intervention achieves the desired selectivity. CoT reduces harmful revision under all-wrong peers substantially (63.9% → 36.5%), but consistently suppresses beneficial revision across mixed and all-correct conditions ($p<.001$). Follow-up analysis suggests why: CoT makes models anchor on their own visible reasoning traces, so when the initial reasoning supports a wrong answer, the model resists correction even when all peers endorse the gold answer—under all-correct peers, staying with the initial wrong answer rises from 27.4% to 50.3% with CoT.

Reflect-then-revise fails differently: it acts as a partial reset toward the Round-1 answer rather than a verifier. It lowers harmful revision under all-wrong peers (63.9% → 41.2%) but also lowers beneficial revision under all-correct peers (52.7% → 33.4%). Reversion analysis confirms the mechanism: harmful Round-2 revisions are reverted at Round 3 more often than beneficial ones (47.9% vs. 29.8% aggregated), consistent with prior findings that LLMs cannot reliably self-correct without external feedback.

Limitations

The authors identify three constraints on generality. First, the multiple-choice format enables unambiguous harmful/beneficial definitions but leaves open-ended generation untested; extending the decomposition would require answer-equivalence judgments with their own biases. Second, simulated peer messages lack iterative dialogue, peer-generated rationales, and emergent multi-agent dependencies, so the patterns may not transfer to live MAS interactions. Third, self-reported 1–10 confidence scores are treated only as auxiliary directional signals, not calibrated probabilities. Additionally, the regression analyses operate at the raw-instance level (~$10^5$ rows), where statistical significance is essentially guaranteed, so the reported $p$-values should be read alongside effect sizes rather than in isolation.

Conclusion

This study moves beyond demonstrating that LLMs conform by directly comparing the two directions of revision within a single factorial design. Its central result—that unanimous wrong peers mislead far more effectively than unanimous correct peers correct—is supported by a fivefold difference in regression odds ratios and replicated across four models and seven tasks. Authority labels shift answers regardless of correctness, and neither CoT nor reflect-then-revise separates harmful from beneficial revision. The practical implication is that multi-agent systems should treat peer answers as claims to be verified against task evidence, not as votes to be aggregated, and should ground any differential weighting of agents in demonstrated calibration rather than role labels. Open questions include whether these asymmetries persist in live multi-agent dialogue with rationale-bearing peers and whether evidence-grounded verification can achieve the selective filtering that prompt-based interventions could not.