acceptodds
Under review as a conference paper at ICLR 2027

EVESkillDebate: Evidence-Verifiable Multi-Agent Debate with Self-Evolving Agent Reasoning Skill

Abstract

Multi-Agent Debate (MAD) can improve language-model reasoning through iterative proposal, critique, and revision. However, agreement under a fixed protocol does not ensure that decisions are evidence-backed, nor does it convert recurring failures into persistent procedural improvements. To address these limitations, we propose a novel paradigm, EVESkillDebate, which consists of Evidence-Verifiable Multi-Agent Debate and Self-Evolving Agent Reasoning Skill. The debate module aims to construct candidate-specific reasoning paths that bind candidate-level claims to document spans and knowledge-graph edges. A frozen verifier assesses citation validity, evidential support, contradiction, and graph consistency, after which an evidence-aware judge aggregates the debated proposals and verification signals. The evolution module contrasts successful and failed (i.e., higher- and lower-utility) execution trajectories to identify recurring skill-level failures and propose targeted updates to the agents' external inference-time policies while keeping the underlying LLM frozen. Candidate skills are evaluated on subsequent batches and deployed only when they satisfy a predeclared feedback gate; post-deployment regressions trigger rollback. Across five public datasets spanning four domains and three LLM families, EVESkillDebate delivers its strongest and most consistent gains on evidence-grounded ranking tasks while achieving high evidence traceability. Evolved skills win 72.4% of non-tied comparisons and achieve 93.7% non-regression, showing that EVESkillDebate converts trajectory feedback into beneficial skill updates in the MAD systems.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Debate, Verifiable Evidence, Self-Evolving Agent Skill, Agentic Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ytLWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EVESkillDebate integrates evidence-grounded debate with the updating of external natural language agent strategies. Before synthesizing the final output, the system links candidate arguments to document snippets, graph edges, or tool outputs, which are then verified by a validation component. Experiments cover many domains such as recommendation systems, rare disease ranking, mathematics, and education, and involve a variety of foundation models.

Critical strengths
  1. The problem is clearly stated and well motivated. Integrating evidence-grounded multi-agent debate with persistent skill evolution addresses both unsupported reasoning and recurring procedural errors.
  2. An extensive evaluation of five benchmarks and three model families supports the claim of their generalisability. In addition, the contributions of each model are analysed through a large number of ablation experiments.
Critical weaknesses
  1. The respective contributions of verifier tuning and skill evolution to the final results have not been sufficiently disentangled. Although the verifier remains fixed during the evaluation phase, this does not account for how its prompts, scoring rules, and thresholds are originally selected or calibrated. Furthermore, repeatedly updating skills under the guidance of a fixed verifier may lead to the skills becoming overfitted to the verifier's specific scoring preferences. The authors could employ an independent verifier or human evaluation to assess the final generated skills.
  2. The additional benefits provided by the debate module are limited, and it remains unclear whether these are sufficient to justify the extra computational cost. In Table 3 on page 8, the debate function yielded improvements of only 0.7 and 0.3% with GPT-5.4-Nano, respectively. As debate often entails additional communication and inference costs, such minimal gains are insufficient.
  3. Many baseline models are variants of MAD. Constructing a robust "single-agent retrieval-plus-verification" baseline will help clarify the marginal value provided by MAD within the constraints of evidence verification.
Questions for authors
  1. How were the verifier's prompts, scoring rules, and thresholds selected and calibrated? Do the evolved skills retain their improvements when assessed by an independent verifier or human evaluators?
  2. Do the debate module's performance gains justify the additional computational cost relative to the no-debate variant?
Reviewer_K9PbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EVESkillDebate, a multi-agent reasoning framework that integrates evidence-grounded debate with self-evolving inference-time policies. The framework constructs candidate-specific reasoning paths using document and knowledge-graph evidence, employs a frozen verifier to assess evidential support and contradictions, and updates external agent skills through contrastive analysis of successful and unsuccessful trajectories. Skill updates are evaluated prospectively under a label-delayed protocol with deployment and rollback mechanisms. Experiments across five benchmarks and three LLM families show substantial improvements on evidence-grounded ranking tasks, while gains on single-answer reasoning tasks are less consistent.

Critical strengths

[#1] A well-structured integration of evidence verification and procedural adaptation. The paper addresses two meaningful limitations of conventional multi-agent debate: the absence of independently inspectable evidence and the inability to retain procedural improvements across instances. The combination of candidate-level evidence paths, explicit verification signals, localized skill edits, and version-controlled deployment provides a coherent framework for improving the accountability and adaptability of multi-agent reasoning.

[#2] Broad empirical evaluation with useful component-level and reliability analyses. The experiments cover five benchmarks, three model families, and a substantial collection of multi-agent baselines. The paper also includes component ablations, computational-cost comparisons, hyperparameter sensitivity, and a human evaluation of 450 claim–evidence pairs. The label-delayed evaluation protocol is particularly valuable because it avoids retrospectively evaluating earlier examples using skills learned from their labels.

Critical weaknesses

[#1] The claimed evidence verifiability is stronger than the demonstrated semantic reliability. Several prominent verification metrics measure structural compliance rather than whether the cited evidence actually establishes the corresponding claim. Although provenance validity, candidate coverage, and graph-path validity reach 100%, these properties are largely enforced by construction. More importantly, the human evaluation finds that only 76.0% of claim–evidence pairs are supported, while 21.1% have insufficient evidence. This creates a substantial gap between the framework's notion of verifiable reasoning and its demonstrated ability to distinguish genuinely supported conclusions from plausible but inadequately grounded ones. The verifier's AUROC of 0.792 is encouraging, but does not establish that the final decision mechanism is robust to semantic verification errors.

[#2] The experiments do not sufficiently isolate the causal contribution and generalizability of self-evolving reasoning skills. The paper reports favorable evolved-versus-unevolved comparisons and component ablations, but its central claim is stronger: that contrastive trajectory analysis identifies reusable procedural failures and produces meaningful, transferable improvements. The reported 72.4% win rate excludes ties, while the 93.7% non-regression rate includes them; these statistics therefore describe different outcome populations. Furthermore, recurring error categories, clause diversity, and cross-setting pattern similarity do not by themselves demonstrate that learned procedural knowledge transfers to genuinely unseen tasks. Since the largest improvements occur on evidence-grounded ranking benchmarks, it remains unclear how much of the overall benefit comes from the verifier and task-specific evidence structure rather than the proposed skill-evolution mechanism itself.

Questions for authors

Q1. Does evidence verification establish semantic correctness, or primarily enforce structural traceability?

Section 4.3 reports perfect or near-perfect scores on several verification metrics, but the human audit shows that 21.1% of claim–evidence pairs lack sufficient support. This discrepancy is important because the paper positions evidence verifiability as a central contribution.

  • How frequently does the final judge select a candidate whose decisive supporting claim is classified as insufficient or contradicted by human annotators?
  • Can the authors distinguish citation validity, evidence relevance, semantic entailment, and answer correctness in a more explicit evaluation? In particular, a valid citation or graph path does not necessarily entail the candidate-level claim.
  • How sensitive is final task performance to verifier errors? A controlled experiment could replace verifier scores with human annotations on an audited subset, or deliberately perturb support and contradiction scores to quantify their downstream effects.
  • Does the evidence-aware judge actually reject unsupported conclusions, or can sufficiently strong agreement among agents override weak evidence? Please report answer accuracy and evidence faithfulness separately for cases with strong support, insufficient support, and contradictory evidence.

These analyses would clarify whether the framework genuinely prevents unsupported reasoning or primarily makes such reasoning easier to inspect.

Q2. What evidence demonstrates that skill evolution learns transferable reasoning procedures rather than exploiting task-specific feedback?

The prospective, label-delayed protocol is a useful design choice, and the paper already reports evolved-versus-unevolved comparisons, component ablations, and cross-task reuse statistics. However, these results do not fully establish the mechanism underlying the claimed procedural improvement.

  • Can the authors report the complete paired outcome distribution—improvements, ties, and regressions—alongside the mean paired utility difference and uncertainty estimates that account for repeated instances or shared experimental settings? The 72.4% non-tied win rate and 93.7% non-regression rate are difficult to interpret jointly without the underlying counts.
  • How does contrastive, clause-localized evolution compare with simpler alternatives under identical evidence, verifier, and inference budgets: generic prompt refinement, random valid clause edits, non-localized edits, or selection among a fixed library of procedural prompts? This would isolate whether the proposed failure-localization mechanism is necessary.
  • Can the authors evaluate skills learned on one domain and deployed without further target-domain feedback on another domain? Figure 2 reports recurring patterns and cross-task reuse, but genuine transfer should be demonstrated through downstream performance improvements on previously unseen target instances.
  • How do skill improvements evolve over time? Please report cumulative online regret or utility relative to a frozen initial policy, together with promotion frequency, rejected candidates, rollback frequency, and performance immediately before and after deployment. Since the deployment gates are empirical rather than statistical guarantees, this would also help establish whether the reported non-regression behavior is robust rather than a consequence of conservative selection or frequent ties.

These experiments are important for substantiating the paper's central claim that EVESkillDebate produces reusable procedural learning, rather than merely improving predictions through a strong evidence-verification pipeline.

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

The paper introduce Evidence-Verifiable Multi-Agent Debate and Self-Evolving Agent Reasoning Skill to improve inference-time multi-agent debate / consistency. It verifies the steps while the agent skill is evolved by contrasting failures and success allowing agents to have access to updated memory/skill at inference time. Authors apply they approach to frontiers and show it improves their performance across five benchmarks.

Critical strengths
  • The approach is straight-forward and doesn't require training so it can be applied to open-source or frontier models.
  • It improves the performance of LLMs from different families on diverse benchmarks.
Critical weaknesses
  • One main issue is that given the nature of method, the comparisons are not matched as the baselines do not have access to task-specific external documents/graphs. So it is not clear how the baselines would compare if they had the same access.
  • Another thing is that it is not clear how this approach work on smaller models.
  • Details are missing for the 450 manually assessed claims.
Questions for authors

see weaknesses

Comments on the paper

3 comments
Area_Chair_d3GxOpenReview ↗

Gentle Reminder: ICLR Rebuttal Discussion Deadline – November 18

Dear Reviewers,

Thank you very much for your time, thoughtful reviews, and valuable contributions to the evaluation process. Thank you to the reviewers who have already interacted and provided additional feedback to the authors!

The authors have carefully addressed your comments and concerns, and the discussion period will conclude on November 18. We would greatly appreciate it if you could take a few moments to review their responses and results, engage in the discussion, and indicate whether their clarifications or additional evidence affect your assessment.

Your continued engagement is essential to ensuring a rigorous, constructive, and fair evaluation process.

Thank you again for your time, expertise, and dedication!

Best regards, Area Chair

Dear Area Chair

We would like to provide a concise update on the discussion and revised submission:

  • Reviewer ytLW: The reviewer has reviewed our response, acknowledged that the concerns regarding generalization and performance gains have been reasonably addressed, concluded that “the paper is now at an acceptable level for multi-agent debate,” and updated the rating to 3. We sincerely appreciate the reviewer’s careful reconsideration.

  • Reviewer K9Pb: There is a clear discrepancy between the reviewer’s updated written assessment and final numerical rating. The reviewer states that the new results “substantially strengthen the paper,” acknowledges that the rebuttal has addressed several major concerns, and indicates an inclination to raise the assessment to Borderline. However, the rating remains at 2 primarily because the available scale does not include a Borderline option. We understand that the ICLR reviewer guide encourages final recommendations to reflect evidence introduced during the discussion. In this spirit, we would be sincerely grateful if you could kindly consider the reviewer’s substantially more positive updated assessment and explicit acknowledgement that several major concerns have been addressed.

  • Reviewer bexf: We fully appreciate that the reviewer may have competing commitments. However, despite several courteous follow-ups, we have not yet received feedback on our response. We would be deeply grateful if you could kindly review our responses and additional evidence at your convenience.

  • Manuscript revision: We have incorporated the additional experiments and clarifications into the revised manuscript, with all changes highlighted in blue.

We sincerely appreciate your valuable time, careful consideration, and support.

Best regards,
The Authors

Dear Area Chair,

Thank you very much for your thoughtful message and for encouraging the reviewers to continue engaging in the discussion. We sincerely appreciate your time, support, and commitment to ensuring a rigorous, constructive, and fair evaluation process.

We are deeply grateful for your assistance and remain available to address any further questions or concerns.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.