acceptodds
Under review as a conference paper at ICLR 2027

Interaction Destroys the Diagnosticity of Agreement Where Truth is Scarce

Abstract

While multi-agent consensus is frequently used as an epistemic proxy for reliability, agreement justifies reliance only when derived from independent evidence. We find that peer interaction compromises this decision-relevant independence, triggering cascades that decouple accuracy from diagnosticity. When three heterogeneous agents deliberate on questions where their independent baseline answers disagree, the resulting two-of-three majorities are false 74.8% (86/115) of the time, versus 23.1% (31/134) for independent votes (Fisher ). A registered blind-revision control separates visibility from revision: private revision keeps conditional falsity inside its registered bound, whereas visibility manufactures 14.2 percentage points [8.3-20.7] more false agreed answers per question on MMLU-hard (McNemar ; replicated at points), without a detected worsening of the majorities that form. We formalize this as the Diagnosticity-of-Agreement Law: agreement signals truth only to the extent agents could have agreed without interacting. Replications indicate that this collapse, which arrives predominantly at the first exchange (62.9% of the increase), tracks a panel's operative ability to verify. A weak panel collapses on GSM8K (80.0% vs. 21.2%) and MMLU (84.3% vs. 57.4%); a mid-capability panel resists on GSM8K ( point) but collapses on MMLU-hard (44.8% vs. 28.9%). Crucially, although deliberation is at least as accurate as self-consistency at a matched number of completions (34.7% vs. 31.0%), it severely degrades diagnosticity (79.6% vs. 46.2% conditional false agreement), indicating that rising accuracy does not by itself warrant reliance on agreement. Finally, a predicted Coverage Law () bounds mitigation on panels that appear able to verify retrieved evidence; on a weaker panel it failed, and checkable evidence restored the bound (4.0% and 2.7% residual).

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Large Language Models, Multi-Agent Deliberation, False Consensus, Epistemic Reliability, AI Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pVVNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether agreement among LLM agents remains a reliable signal after interaction. The main finding is that deliberation can improve overall accuracy while substantially increasing the error rate among formed majorities. The paper attributes this to interaction-induced dependence and false consensus, and further studies truth scarcity, peer injection, grounding, and verification as possible mechanisms and boundary conditions.

Critical strengths

The paper asks an interesting question that is different from simply measuring whether multi-agent debate improves accuracy. The distinction between overall performance and the evidential value of agreement is useful. The paper also includes a relatively extensive set of controls, replications, and intervention experiments.

Critical weaknesses
  1. The paper seems to conflate the absence of an independently correct answer with the absence of useful distributed information. A truth-scarce case (k=0k=0) does not imply that the agents collectively lack information about the truth. Different agents may hold complementary partial evidence that becomes useful only after communication. This matters especially because the agents here appear to share a cooperative objective, unlike classical information-elicitation settings with strategic human participants. The current experiments mainly show a failure of unstructured answer-sharing, but the broader claim that interaction cannot create diagnostic agreement seems too strong.

  2. The novelty over recent work on false consensus, diversity collapse, and debate-vs-voting is not yet fully convincing. The paper itself discusses closely related work on manufactured overconfidence, false consensus, confidence/diversity effects, and debate versus voting. The main distinction appears to be the specific notion of “diagnosticity of agreement” and the associated controls. I would like a clearer explanation of which empirical or theoretical conclusions would not already follow from these prior results.

  3. Some of the main empirical conclusions depend heavily on conditioning and pool construction. The headline 23.1% versus 74.8% comparison conditions on non-unanimous majorities, but deliberation substantially changes which questions enter that set. In other words, part of the effect is a composition shift caused by manufactured agreement. The paper analyzes this mechanism, but the framing sometimes reads as if deliberation directly makes agreement intrinsically less reliable, rather than changing the process by which agreement is formed. In addition, several results depend on divergent-start or truth-scarce subsets, so the deployment-level interpretation should be stated more cautiously.

Questions for authors
  1. How would the main claim change if agents hold complementary partial information that is individually insufficient but jointly sufficient to recover the correct answer? Would such a setting violate the proposed Diagnosticity-of-Agreement Law?

  2. What is the strongest empirical distinction between this work and prior work on false consensus, diversity collapse, and debate versus voting? Which of the paper’s main conclusions would not be obtained from those earlier findings?

  3. Since deliberation changes the set of questions that form a majority, how much of the diagnosticity gap remains under a comparison that holds the agreement set fixed, rather than conditioning on different post-treatment agreement events?

Reviewer_TFvwOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether the consensus reached by multiple LLM agents after interacting and discussing with each other can still be considered a reliable indicator of answer correctness. By comparing independent voting with multi-round deliberation across different model capabilities and task types, the authors find that when agents lack sufficient knowledge, interaction can lead to the spread of incorrect information and eventually result in false consensus. The experimental results also shows that while multi-round deliberation may improve overall accuracy, it can actually reduce how reliably agreement indicates correctness. To further explain this phenomenon, the authors also propose the Diagnosticity-of-Agreement Law, the Anchor Principle and the Coverage Law. They investigate the underlying mechanisms and conditions through false information injection, blind revision controls, pre-registered replications, and retrieval-augmented experiments. The results suggest that false consensus is closely related to agents' independent knowledge and their ability to verify information, while providing verifiable external evidence can help reduce this problem under certain conditions.

So, I think the main contribution of this paper is that it distinguishes between overall accuracy and consensus reliability in multi-agent systems. This provides empirical evidence for understanding like how errors spread during multi-agent interactions and may help guide the design of more reliable collaboration mechanisms.

Critical strengths
  1. The paper explores an interesting direction by distinguishing between accuracy and consensus reliability in multi-agent systems. Most of the previous studies I have read focus on whether multi-agent discussion, debate or like collaboration can improve accuracy. This paper goes a step further by asking whether the consensus reached after discussion is actually reliable. I think one interesting finding is that multi-round deliberation can improve overall accuracy while making consensus a less reliable indicator of correctness. This highlights an important issue that is often overlooked in multi agent sysytem research.

  2. The experiments are quite comprehensive and provide a deeper analysis of how false consensus forms. The authors not only compare independent voting with multi-round deliberation but also use Blind Revision, false information injection, comparisons across different model capabilities, and pre-registered replications to validate their findings. I also noticed that they use a human-validated evaluator and conduct experiments without relying on an LLM Judge. In addition, the appendix reports hypotheses that were not supported by the experiments, which makes the evaluation more transparent.

Critical weaknesses
  1. The paper mainly focuses on a setup with three agents and four rounds of fully connected deliberation. Although the authors also include additional experiments with five agents, they do not evaluate more structured approaches such as Role-based Debate or Hierarchical Collaboration. However, the general form of the Diagnosticity-of-Agreement Law covers a broader range of interaction protocols. I think current experiments are not enough to show whether similar drops in consensus reliability would also occur under other collaboration mechanisms.

  2. In Section 3.1 the authors mainly focus on questions where agents initially disagree and evaluate non-unanimous majorities. The reported 23.1% (31/134) and 74.8% (86/115) are actually calculated from different sets of agreement events. Although the paper provides paired analyses and unconditional error rates, I think it is still unclear how well this conditional comparison reflects consensus reliability in real-world task distributions without such filtering.

  3. The Coverage Law relies on the law of total probability and several specific assumptions. However, in Section 3.4 the observed false consensus rates on weaker models reach 34.0% and 26.7% at retrieval coverage levels of 90% and 95%, clearly exceeding the predicted bounds. The authors then introduce Operative Verification Capability to explain these results, but this capability is not independently measured. I think this makes the conditions under which the Coverage Law holds less clear and somewhat weakens its theoretical claims.

Questions for authors
  1. Does the current evidence support generalizing the Diagnosticity-of-Agreement Law beyond Fully-connected Deliberation to structured approaches such as Role-based Debate or Hierarchical Collaboration?

  2. In Section 3.1 the reported 23.1% and 74.8% come from different agreement events on a filtered dataset. Would the drop in consensus reliability remain significant on an unfiltered task distribution?

  3. In Section 3.4 the Coverage Law fails on weaker models. Why should this be viewed as a limitation of its applicability rather than a failure of the prediction? Is there an independent measure of Operative Verification Capability to support this explanation?

Reviewer_9fwxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies the effect of deliberation on the quality of collective decision making in multi-agent LLM systems. Motivated by the independence assumption in epistemic theory, the paper claims that agents’ deliberation imposes harm to collective decision making quality by introducing correlation. The paper reports empirical results based on three LLMs and multiple benchmarks, showing evidence supporting its configurations. Moreover, the paper summarises the findings in three principles, i.e., the diagnosticity-of-agreement law, the anchor principle, and the coverage law.

Critical strengths
  1. The work studies an important question for multi-agent LLM systems. The motivation is also rooted in classic epistemic theory: decision independence plays an important role in the quality of collective decision-making events.
  2. The empirical study provides interesting information to support the claims.
Critical weaknesses
  1. Clarity: It is hard to follow the paper because of its unclear presentation. The paper uses mixed terms to describe the summarised laws, empirical results, and discussions, causing confusion here and there. In the empirical section (Section 3), the information density is high, where descriptions use a mixture of intuitive descriptions, statistical results, and unclear abbreviations to convey ideas. It is hard to understand the main results without a good arrangement of content. The insufficient introduction of the terms, abbreviations, and special symbols (e.g., the benchmarks) also makes the main text not self-contained.
  2. The work uses a small number of LLM agents in the experiments, limiting the scope of the empirical observation. The Condorcet theorem shows that the quality of the collective decision making improves as the number of informed agents increases. Therefore, agent number is a critical parameter in the scenario. I would recommend the work adds more experiments to reveal the influence of agent number.
Questions for authors

See the weakness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.