acceptodds
Under review as a conference paper at ICLR 2027

Breaking the Martingale Curse: Multi-Agent Debate via Epistemic Asymmetry

15 views

Abstract

Multi-Agent Debate (MAD) is a widely used approach to improving the reasoning of large language models (LLMs), but it can fail when agents make correlated errors and converge on the same wrong answer. Under symmetric, linear belief aggregation, the aggregate belief in the correct answer is a martingale: its expectation does not change across rounds, so debate cannot recover from an incorrect initial majority. We refer to this limitation as the Martingale Curse. We propose AceMAD, a protocol that exploits an epistemic asymmetry between two kinds of agents. Anchored Agents follow the base model's dominant reasoning path and tend to assume that their peers share their view; Exploratory Agents are perturbed away from that path and, when they recognize the majority's error, can anticipate it. AceMAD elicits this second-order information through peer prediction: every agent also predicts the average belief of its peers, the prediction is scored with a Brier-type proper score (the ACE Score), and agents with better-calibrated predictions receive multiplicatively larger influence (ACE Reweighting). We prove that, when Exploratory Agents maintain a persistent score advantage (an ACE Gap), the aggregate belief becomes a submartingale with positive drift toward the correct answer. On challenging subsets of six benchmarks, AceMAD recovers correct minority answers under correlated errors and outperforms majority voting and existing MAD protocols.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Debate, Martingale Curse, Large Language Model Reasoning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_4WJ3OpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors propose AceMAD to mitigate the Martingale Curse, which is a known problem for multi-agent debate. The AceMAD protocol attempts to increase the influence of agents whose predictions for their peers' beliefs are more accurate. Empirical results on a targeted subset of datasets show clear advantage of AceMAD over baselines.

Critical strengths
  1. The paper is well-written and well explains the complicated theoretical intuition behind the proposed method.
  2. The motivation and the way the authors try to solve the Martingale's Curse are highly relevant to the core problem commonly observed in multi-agent debate.
  3. The authors provided a broad range of experiments, and the "challenging" subset is well targeted for the problem setup.
  4. The experimental results are strong and well support the work's method.
Critical weaknesses
  1. My main concern is related to the soundness of the theorems that support the method. Specifically, it feels like the assumptions stated in Theorem 4.6 can be a bit strong. In particular, the persistent ACE Gap assumption requires the exploratory agent to maintain a positive expected peer-prediction score advantage over the anchored majority, conditional on the debate history at every round. This seems potentially restrictive, especially early in debate when little interaction history is available. Could the authors report the exploratory and anchored agents' peer-prediction errors and score differences throughout the debate, including how frequently the exploratory agent lacks this advantage?

  2. Another concern about the theory is that it seems the theorems don't fully account for the agents' changing beliefs throughout the debate. I've checked the derivations for Theorem 4.6, and to my understanding, the change in aggregate belief is only expressed by the changes in influence weights, and omits the effect of the changes in the agents' own beliefs. Since agents revise their beliefs during debate, these additional terms could offset the benefit of re-weighting even when the exploratory agent retains a positive signal gap. Could the authors account for these terms in the proof or explicitly state the additional assumptions needed? Also, it is unclear whether the calibration advantage persists as agents revise their beliefs and the initially minority answer becomes more widely supported. While Remark 4.7 discusses the minority gaining majority influence, it does not establish the subsequent evolution of beliefs and ACE scores. Clarification and empirical trajectories would help assess the mechanism's behavior through this transition.

  3. The nature of this work and its underlying theorems work on expected values. However, the paper mainly reports a single-run result, if I am not wrong. Could the authors clarify the number of independent runs and report means and uncertainty estimates across repeated runs? It would also be helpful to include a small empirical analysis of the quantities underlying Theorem 4.6: e.g., the exploratory agent’s ACE score advantage over the anchored majority, the corresponding signal gap in probability assigned to the correct answer, the evolution of influence shares, and the change in aggregate belief in the correct answer across rounds. For example, estimating the score-gap and signal-gap parameters (gamma) and (beta), and reporting how often these gaps diminish or reverse, would help assess whether the theorem's assumptions describe the observed debate dynamics.

  4. The results in Table 1 and Table 2 are at odds. To my understanding, both tables report the baselines' accuracy on the same subset of data with the same backbone LLM, GPT4o-mini. Could the authors clarify why the values are different in the two tables?

Questions for authors

See Weaknesses

1 comment

Thank you for the authors' effort in conducting the extensive amount of additional experiments.The additional details are also helpful. It is also promising that the authors have updated the manuscript accordingly. I believe all my concerns have been sufficiently addressed, and would like to strongly advocate for acceptance. I am increasing my score to 4.

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

This paper proposes AceMAD to address cases where multi-agent debate fails because most agents make similar mistakes and overlook the correct minority answer. It asks each agent to predict its peers' answer probabilities, scores these predictions using a Brier-type score, and gives more weight to agents with more accurate predictions in the final aggregation.

Critical strengths
  1. I think the focus on correlated errors is well motivated, as this is a setting where majority voting can easily fail.

  2. Using peer prediction instead of self-reported confidence to weight agents is an interesting and intuitive idea.

Critical weaknesses
  1. ACE relies on fine-grained probabilities directly reported by LLMs, but the paper does not examine how stable or reliable these estimates are. Small variations in the reported probabilities could affect ACE scores and the resulting agent weights. It is also unclear whether such precise probabilities are necessary, or whether a simpler rank-based approach would be more robust.

  2. I noticed some differences in the baseline results between Tables 1 and 2. For example, majority voting reports 14.0% and 18.96%, while Decentralized MAD reports 22.1% and 28.62%. Since Section 6 mentions using the same subsets and baselines as Section 3, it would be helpful to clarify whether there were any changes in the experimental settings or if these differences come from separate runs.

Questions for authors
  1. How stable are ACE scores when probabilities are elicited repeatedly from the same debate history? A comparison with rank-based scoring and an analysis of whether higher ACE scores correspond to more accurate answers would also be useful.

  2. In the 4+1 setup, the exploratory agent predicts four similar anchored agents, while each anchored agent also has to predict the exploratory one. Could this difference in prediction difficulty partly explain the ACE gap? Comparing agents on the same peer group might help clarify this.

Reviewer_m4j9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AceMAD, a method for improving multi-agent debate (MAD) in large language models by leveraging cognitive asymmetry. To address the problem of incorrect consensus arising from correlated errors among agents, the authors divide agents into Anchored Agents and Exploratory Agents and require each agent to predict the average beliefs of its peers. The method then uses the ACE Score to evaluate prediction accuracy and adjusts each agent's influence on the final answer through a multiplicative weight-update mechanism.

Theoretically, the paper analyzes the relationship between disparities in prediction scores and positive drift in aggregated beliefs. Experiments on challenging subsets of six benchmarks show that AceMAD outperforms majority voting and several existing debate methods across multiple models and tasks.

Critical strengths
  1. The paper addresses the problem of incorrect consensus in multi-agent debate, particularly how to distinguish valuable minority opinions from dominant but erroneous views. This is a meaningful research direction.

  2. Evaluating agents based on their ability to predict peers' belief distributions, rather than relying solely on self-reported confidence or majority voting, offers a novel approach to weighting agents without external labels.

  3. The paper attempts to theoretically connect advantages in peer-prediction scores with changes in influence weights and positive drift in aggregated beliefs, moving beyond a purely empirical design.

  4. The authors evaluate the method across six benchmarks and multiple model families, including comparisons with different debate protocols, sources of heterogeneity, and ablation settings. These experiments provide preliminary evidence of the method's effectiveness.

Critical weaknesses
  1. The key contribution of AceMAD lies in introducing the ACE Score, which measures the accuracy of predicting other agents' belief distributions rather than answer correctness. However, more accurate predictions of group opinions do not necessarily imply more reliable answers. The relationship between these two factors requires more thorough empirical validation.

  2. Theorem 4.6 relies on assumptions including a persistent ACE score advantage, a higher belief in the correct answer among Exploratory Agents, and non-degenerate weights. The paper does not sufficiently establish how generally these conditions hold in practical debates, leaving a gap between the theoretical guarantees and the observed empirical improvements.

  3. The proportion of Exploratory Agents, temperature, prompting strategy, reweighting coefficient, and number of debate rounds may substantially affect performance. The current ablation studies do not fully characterize the method's sensitivity to these hyperparameters. Moreover, the impact of heterogeneity between Anchored and Exploratory Agents on overall MAD performance remains unclear.

  4. Increasing the number of debate rounds T or agents N may incur substantial token costs and increase coordination complexity. Although Figure 4(b) examines the scaling effect of N, the paper does not report the associated computational time overhead.

Questions for authors
  1. How can the authors verify that the group-prediction ability measured by the ACE Score is consistently associated with final answer correctness? When most agents strongly agree on an incorrect answer, could the proposed mechanism still reward the resulting erroneous consensus?

  2. Could the authors report results on the full test sets and across different difficulty levels? In addition, does selecting challenging subsets introduce potential selection bias?

  3. How does the method perform when Exploratory Agents fail to identify the correct answer or do not exhibit superior belief-prediction ability compared with other agents? Could the authors provide a more systematic analysis of failure cases?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.