acceptodds
Under review as a conference paper at ICLR 2027

Reasoning Jury: Multi-Model Consensus for Evaluating Reasoning Traces

Abstract

Improving reasoning LLMs requires the ability to judge the quality of long reasoning traces for effective reasoning data curation, strong training signals during reinforcement learning, and an in-depth understanding of reasoning behaviors during model performance evaluation. Additionally, surfacing reasoning mistakes that the model makes would enable improving the model's performance at runtime through providing feedback. Due to the difficulty of this complex task on long reasoning traces, single-model judges (even frontier models) do not do well at identifying reasoning defects. Additionally, leveraging frontier models during online training of reasoning LLMs is generally prohibited due to guardrails in terms of use. In this work, we introduce Reasoning Jury, a system that replaces the single judge with a jury of LLMs and a moderated consensus mechanism, to improve the fidelity of judgments for identifying reasoning defects. The moderator derives a consensus through deliberation amongst jurors or consolidation of judgements. We show that Reasoning Jury with a jury of open-weight models is able to significantly outperform frontier models (Opus, Sonnet, and Gemini) at correctly identifying reasoning defects. Besides accuracy performance improvements, the aggregated cost of the jury (initial verdicts, deliberations, consolidation, etc.) is a fraction (8 to 16%) of the cost of running frontier models in LLM-as-a-judge setup. We also show how these judgements can be leveraged to understand failure modes of reasoning LLMs on benchmarks, which allows much deeper understanding of a model's performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Reasoning LLM Evaluation, LLM-as-a-judge, Deliberation

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_Xtz1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents Reasoning Jury, which uses multiple LLMs to identify errors in reasoning traces and combines their findings through consolidation or moderated discussion. Each finding includes the affected steps, a diagnosis, severity, and evidence. Experiments evaluate error localization on Hard2Verify and DeltaBench, token-priced cost, and the usefulness of the feedback for retrying AIME2026 problems.

Critical strengths

The paper addresses a useful problem, and the gains from combining inexpensive judges are promising. The comparisons with simple aggregation help explain where those gains come from. The AIME experiment also provides useful evidence: rich feedback improves retry accuracy from 71.2% to 76.2% over step locations alone, with an analysis that accounts for dependence among samples from the same problem.

Critical weaknesses

1. The methodological contribution is limited

Multi-model judging and multi-agent debate are already established [1, 2]. Here, the main change is to apply them to step-level reasoning defects with structured reports. I do not see a substantial methodological advance, particularly when simple union captures most of the localization gains and slightly outperforms deliberation on DeltaBench. The paper shows that collecting more judgments helps, but does not make a convincing case for its particular consensus mechanism.

2. The richer judgments are not adequately validated

The paper's case for going beyond simple aggregation rests on better diagnoses, severity ratings, and supporting evidence. Yet the main benchmarks only score error locations. The retry experiment is useful, but some feedback supplies explicit corrections, and better final answers do not establish that the underlying diagnoses are reliable. Without directly checking these judgments or comparing equally rich feedback from simpler alternatives, I cannot tell whether consensus actually produces a better evaluator.

3. The numerical inconsistencies weaken my trust in the results

Appendix L acknowledges unresolved inconsistencies in the Hard2Verify metrics and conflicting entries across tables. I appreciate the disclosure, but these are still the results used to support the paper's conclusions. Until they are reconciled, I cannot confidently rely on the reported comparisons. This materially affects my assessment of the paper's credibility.

[1] Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models.

[2] Improving Factuality and Reasoning in Language Models through Multiagent Debate.

Questions for authors

My main questions follow directly from the concerns above: what does the consensus mechanism add over simpler aggregation, and can the authors reconcile the reported results and establish the reliability of the resulting judgments?

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

This paper aims to develop a multi-jury (evaluator) framework for evaluating reasoning traces and finding the problematic reasoning steps. The main idea here is that instead of using one jury to audit a reasoning rollout, they ask multiple judges to independently audit the rollout. Then, they introduce two mechanisms to combine the independent results. The first one is referred to as consolidation, where a single jury gets all reports from the previous step and is asked to come to a conclusion. The other one is called deliberation, in which an orchestrator is asked to help the jury communicate with each other and come to a conclusion without adding any judgement on its own. At the end, the authors showed in setups such as Hard2Verify, we may get the same result when using a jury of small open-source models vs. using a single frontier model while incurring much less expense.

Critical strengths
  1. I think the paper is generally clear and easy to follow. In particular, Figures 1 and 3 do a good job of explaining the main concepts. I also think the main message of the paper (better performance of open source jury vs a single frontier model) is interesting, although that is not the case for gpt-5.4.
Critical weaknesses
  1. I think the main novelty of the work should be in the deliberation method since consolidation is just two different stages of prompting. However, the problem is that for most cases the performance of consolidation and deliberation is close, and I expect the authors to address this. I think in general there must be a deeper analysis on why using deliberation. For example, some examples of the communication between models.

  2. As mentioned by authors in their limitation section, one simple baseline in general is to ask a model to retry without giving the model the defect in their reasoning, while the paper is missing that baseline. In general, the paper has limited experiments for the application of this designed system of juries for improving model accuracy.

Questions for authors
  1. please look at the weaknesses.

  2. How did the authors come up with the prompts for the jury and mediator? It would be much more interesting if the authors had done some sort of optimization on the prompt there on a training set.

  3. Could authors add more experiments for section 3.3? Again, in their only application-based experiment, the authors are missing several baselines, including blind reply. Also, I think having an experiment on a harder problem than AIME would be beneficial.

  4. Why did in the section 3.3 experiment (Table 3) some rollouts go from being right to wrong after getting the guidance?

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

This paper introduces Reasoning Jury, a multi-model framework for detecting and diagnosing errors in long reasoning traces. The framework consists of two stages: independent judgments from multiple LLM jurors, followed by consensus formation through either one-shot consolidation or moderator-guided deliberation. Each identified defect is associated with specific reasoning steps, an explanation, a severity level, and supporting evidence. Experiments on Hard2Verify and DeltaBench demonstrate that multi-model juries substantially outperform their individual constituent judges. For example, a homogeneous jury of three gpt-oss-120b instances achieves 82.3 Balanced F1 on Hard2Verify, compared with 73.7 for Claude Opus 4.6, at approximately 15.6% of its estimated token-priced cost. The paper further demonstrates that providing structured defect feedback improves retry accuracy on AIME2026 from 71.2% with step locations alone to 76.2% with additional diagnoses and severity information.

Critical strengths
  1. Strong empirical improvements with practical cost advantages.

The paper demonstrates substantial improvements in reasoning defect detection by combining judgments from multiple models. Particularly encouraging is the observation that homogeneous juries of relatively inexpensive open-weight models can outperform substantially stronger individual judges. The experiments cover two benchmarks and investigate jury composition, moderator selection, and aggregation strategies. These results suggest a potentially cost-effective approach to scaling reasoning verification, which is particularly relevant for large-scale training data curation.

  1. Actionable error localization and downstream validation.

Beyond producing binary correctness judgments, the framework generates structured, step-grounded feedback describing reasoning defects. The AIME2026 retry experiment provides useful evidence that such feedback can improve downstream reasoning performance. The improvement from 71.2% to 76.2% on flagged traces, supported by a positive problem-level confidence interval, demonstrates the potential utility of detailed reasoning feedback beyond error localization alone.

Critical weaknesses
  1. Unresolved numerical inconsistencies undermine confidence in the experimental results.

Appendix L explicitly reports that 35 of 74 full-metric rows cannot be reconciled with consistent integer confusion counts under the assumed Hard2Verify scoring universe. It also identifies specific inconsistencies between reported precision, recall, and F1 values. The original record-level predictions and run logs are not included in the provided source bundle.

Although the authors acknowledge that this consistency check depends on the assumed scoring configuration, these unresolved discrepancies raise concerns about the reliability and reproducibility of the quantitative results. Given that several comparisons involve relatively small performance differences, resolving these inconsistencies is important before drawing firm conclusions about the effectiveness of the proposed methods.

  1. The incremental contribution of moderated deliberation remains insufficiently established.

While combining multiple independent judgments produces substantial gains over individual judges, the experiments suggest that much of this improvement can be achieved through simple aggregation.

In Appendix H, union aggregation matches or exceeds deliberation in two of the six Hard2Verify jury configurations. On DeltaBench, union aggregation slightly outperforms deliberation in both Balanced F1 (61.7 vs. 61.5) and macro-F1 (48.2 vs. 47.9). This raises questions about whether the additional complexity and computational overhead of moderated deliberation are sufficiently justified.

Furthermore, the main benchmark evaluations reduce structured defect reports to binary step-level labels, without directly evaluating the correctness of diagnoses, severity assignments, or supporting evidence. Although the retry experiment provides encouraging downstream evidence, it lacks a blind-retry baseline and combines diagnosis, severity, and occasionally explicit corrections in a single condition.

Consequently, the results establish the benefits of aggregating multiple judgments more convincingly than the specific benefits of the proposed deliberation mechanism and its richer structured outputs.

Questions for authors
  1. Could the authors clarify the exact scoring universe and evaluation procedure underlying the inconsistencies identified in Appendix L? Can the reported results be regenerated from record-level predictions, and do the main conclusions remain unchanged after reconciliation?

  2. Can the authors provide paired statistical comparisons between deliberation and simple union aggregation, ideally under comparable computational budgets? This would help establish whether deliberation provides meaningful improvements beyond pooling independent judgments.

  3. Can the authors provide evidence for the correctness of the generated diagnoses and severity labels, for example through human annotation of a representative sample? Additionally, how does the retry improvement compare with a blind-retry baseline, and what are the separate contributions of diagnosis and severity information?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.