Phantom Consensus: Measuring Social Reasoning in Multi-Agent LLM Decision-Making

AI summary

Overview

Research area: Multi-agent large language model (LLM) systems, social reasoning, and benchmark design for collective decision-making.

Technical level: Intermediate. The social-psychology framing is accessible, but the paper assumes familiarity with multi-agent evaluation, accuracy metrics, and belief-elicitation protocols.

Scope: The paper introduces PHANTOM-Bench, a five-condition benchmark of 480 three-way decision scenarios run with four-agent groups across nine LLMs, to measure whether agents publicly converge on decisions that the group's own recorded private judgments do not support.

What This Paper Is About

When several LLM agents discuss the same case, they can end up agreeing on an answer that none of them actually believed when reasoning alone. The authors call this failure phantom consensus: agents read vague public comments from peers, guess what those peers privately think, guess wrong, and then shift their public decision toward that imagined group belief. The paper's goal is to measure this pattern directly — by recording each agent's private judgment first, then comparing it with public decisions and with each agent's stated estimate of what the rest of the group privately decided.

Key Contributions

1. A new benchmark. PHANTOM-Bench measures phantom consensus in controlled four-agent LLM collectives, using 480 scenarios with a balanced three-way action space (proceed, block, defer or review) and a designated gold decision plus a plausible-but-wrong "phantom target."

2. Five belief-context conditions. The design separates isolated private judgment (Private), judgment after ambiguous public signals (Observed Public), recovery after recorded private decisions are revealed (Informed Transparency), a neutral-text control (Neutral Public), and direct visible conformity (Explicit Public Vote).

3. New diagnostics for belief failure. The paper defines metrics for accuracy drop (Private-Public Gap, PPG), collective failure (Collective Rationality Failure Rate, CRFR), phantom-specific collective failure (P-CRFR), phantom-aligned reported beliefs (Phantom Belief Alignment, PBA), second-order belief error (Perceived Belief Error, PBE), agent-level phantom-induced failure (PIFR), and recovery (Transparency Recovery Rate, TRR).

4. Evidence that ambiguity alone can break collective accuracy. Ambiguous public signals — with no stated vote or majority label — can turn scenarios where at least three of four agents were privately correct into public outcomes where at most one remains correct, alongside large errors in agents' reports of what peers privately decided.

Main Findings

• Every model degrades after ambiguous signals. Private Acc falls to a lower Public Acc for all nine models. The largest drops are Qwen-2.5-7B (77.0% to 29.2%, PPG of 47.8 percentage points) and Qwen-2.5-32B (73.1% to 42.1%, PPG of 31.0 percentage points).

• Collective failures, not just individual errors. CRFR — scenarios with at least three of four agents correct privately but at most one correct publicly — reaches 47.3% for Qwen-2.5-7B and 43.3% for Qwen-2.5-32B. Non-trivial rates also appear for Gemini-2.5-Pro (16.5%) and Mistral-Small-24B (22.9%), while Gemini-2.5-Flash is lowest at 9.8%.

• Failures often land on the phantom target. P-CRFR reaches 39.4% for Qwen-2.5-7B and 20.8% for Qwen-2.5-32B, meaning the collective failure is frequently aligned with the specific wrong answer the ambiguous signals were designed to suggest.

• Reported group beliefs are highly inaccurate. PBE, which compares an agent's estimate of the other agents' private majority against their actually recorded private decisions, reaches 77.1% for Gemini-3.5-Flash, 74.1% for Gemini-2.5-Pro, and 73.8% for Llama-3.1-8B. PBA (phantom-aligned reported group beliefs) is 65.8% for Gemini-3.5-Flash and 63.0% for Llama-3.1-8B.

• Belief error and public failure dissociate across models. Llama-3.1-8B has high PBE (73.8%) but comparatively low PIFR (12.9%), whereas Qwen-2.5-7B has lower PBE (55.6%) but much higher PIFR (42.2%); Qwen-2.5-32B has PIFR of 22.5%. The authors note that aggregate metrics cannot establish whether belief errors and action changes co-occur within individual decisions.

• Neutral text is weaker than ambiguous signals, but explicit votes vary by model. Gemini-2.5-Pro shows 19.4% observed-public phantom alignment versus 6.5% under Neutral Public and 15.1% under Explicit Vote (differences of 12.9 and 4.3 percentage points); Mistral-Small-24B shows 22.8% versus 6.1% and 22.0%. Qwen-2.5-7B, however, shows comparable observed-public (59.4%) and explicit-vote (61.0%) alignment.

• Transparency repairs many failures without revealing the answer. Informed Acc returns close to Private Acc for most models: Qwen-2.5-7B rises from 29.2% Public Acc to 77.1% with 99.6% TRR; Qwen-2.5-32B rises from 42.1% to 72.5% with 97.9% TRR; Mistral-Small-24B reaches 100.0% TRR. Recovery is conditional, however — Llama-3.1-8B has 62.3% TRR while its overall accuracy declines from 59.8% Public Acc to 56.1% Informed Acc.

Methodology in Plain English

Each benchmark item is a short institutional decision case — drawn from domains such as financial compliance, content moderation, hiring fairness, healthcare privacy administration, and workplace operational safety — where an agent must choose one of three actions: proceed, block, or defer or review. Every scenario has a fact-supported gold decision and a deliberately plausible wrong alternative (the phantom target), and the dataset is balanced at 160 scenarios per gold action class, with 480 scenarios and four agents each (1,920 agent-level examples per condition).

The pipeline runs the same scenario through four independently instantiated agents under different information conditions. In the Private condition, agents see only the case. In the Observed Public condition, they additionally see four scripted ambiguous comments that hint at a group tendency without ever stating a vote or majority; each agent then gives a public decision and a guess at what the rest of the group privately decided. The Informed Transparency condition then reveals the other agents' recorded private decisions (never the gold label). The two controls — Neutral Public comments with no implied norm, and Explicit Public Vote recommendations — isolate whether shifts come from public text as such or from old-fashioned visible conformity.

Because private judgments are recorded before any social signal, the authors can compare what the group actually privately decided against what agents think the group privately decided. Scenario quality was checked by two annotators on 150 randomly sampled scenarios using a rubric covering gold-label clarity, phantom-target validity, signal ambiguity, private-judgment leakage, explicit majority pressure, norm-source match, and target-direction validity, with Cohen's κ = 0.83 and adjudication of disagreements.

Why This Matters

Impact on research. Most prior work on LLM conformity studies visible peer influence — stated answers, votes, recommendations. This paper targets second-order belief: agents acting on an inference about what peers privately think. It argues that evaluating multi-agent systems by final agreement alone is insufficient; evaluation should ask whether agreement reflects independently measured private judgments or an imagined consensus.

Real-world applications:

• Compliance and financial-crime review, where agents assess transfers, KYC documentation, and novel business models (the paper's worked example is a $5 million transfer case grounded in FinCEN guidance).
• Content moderation, where vague peer commentary could push reviewers toward permissive or over-blocking outcomes.
• Hiring and fairness review, where implied consensus about a candidate could override fact-based evaluation.
• Clinical and administrative privacy decisions (HIPAA-grounded scenarios), where escalation versus approval carries real risk.

Industry relevance. The finding that disclosing recorded private first-pass judgments repairs many failures — without revealing the gold answer — is directly actionable for teams designing multi-agent review pipelines. The authors' practical recommendations are to preserve private first-pass judgments, expose disagreement, and stop treating public statements as reliable evidence of peers' private decisions.

Future Directions

• Isolate the causal mechanism. The authors state that the condition comparisons do not establish a distinct causal mechanism, since conditions differ in comment content and belief elicitation; a design that holds content constant while varying only inferred-norm strength would help.

• Separate belief correction from information gain. Transparency recovery cannot be attributed solely to fixing misperceived group beliefs, because revealing private decisions also supplies task-relevant peer answers. Designing disclosures that inform about beliefs without leaking decision content is an open problem.

• Link belief error to action within individual decisions. The paper notes that aggregate PBA/PBE and PIFR metrics cannot show whether belief errors co-occur with action changes in the same decision; per-decision analysis could test whether misperception drives the shift.

• Explain why models differ. Models vary widely in how often belief-level errors convert into public-stage phantom failures (Llama-3.1-8B versus Qwen-2.5-7B), and the paper does not identify what drives this variation. Relatedly, the provided text reports no compute, cost, or human-baseline comparisons, and the truncation cuts off mid-way through Appendix B.3.

Target Audience

Researchers and engineers building or evaluating multi-agent LLM systems — particularly those designing deliberation, critique, voting, or consensus pipelines — will benefit most. It is also useful for social-behavior researchers studying conformity and pluralistic ignorance in artificial agents, and for practitioners in compliance, moderation, hiring, and healthcare administration who need to know whether an LLM panel's "agreement" reflects independent judgment. Readers should be comfortable with benchmark metrics and multi-condition experimental design; the social-psychology framing (the Abilene paradox, pluralistic ignorance, the spiral of silence) is explained from first principles.

Authors’ abstract

Multi-agent LLM systems increasingly aggregate judgments across model instances, but reasoning about other agents can itself introduce failure. We study a setting in which LLMs shift their public decisions after observing ambiguous social signals while reporting inaccurate estimates of what other agents privately judged, even when the group's recorded private judgments do not support the apparent consensus. We call this failure phantom consensus and study it as a…