acceptodds
Under review as a conference paper at ICLR 2027

Communication as Evidence Routing: Adaptive Relational Graphs via Optimal Transport for Multi-Agent LLM Reasoning

Abstract

Multi-agent large language model (LLM) systems offer a promising paradigm for complex reasoning by integrating complementary expertise across multiple models. However, existing agent communication strategies rely on either free-form interaction or answer-level aggregation, making reasoning susceptible to dominant-answer collapse, where prevailing answers prematurely suppress valuable dissenting evidence. To address this, we reinterpret multi-agent communication as evidence routing and propose Adaptive Relational Graphs via Optimal Transport (ARGOT), a relational graph framework for multi-agent LLM reasoning. ARGOT represents agents’ initial responses as evidence distributions and uses optimal transport to construct a task-adaptive graph comprising support and challenge relations. These typed relations enable conflict-aware message passing, allowing consistent and conflicting evidence to be propagated and integrated differently during reasoning. A trajectory-level verifier then selects the final answer by evaluating routed evidence and reasoning trajectories, rather than relying solely on answer frequency. Across six reasoning benchmarks, ARGOT achieves the highest average performance and ranks first on four tasks. Further analyses show it preserves informative disagreement, improves evidence aggregation, and mitigates premature consensus. These results establish structured evidence routing as an effective, principled approach to more reliable multi-agent LLM reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Graph-based Reasoning, Collaborative LLM 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_fYnoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARGOT, a training-free framework that organizes heterogeneous LLM collaboration as evidence routing. It selects task-relevant agents, converts their initial responses into distributions, and uses optimal-transport costs and self-reported model confidence to construct support relations and protected challenge relations. Two directed communication phases refine responses, after which a trajectory-level verifier compares evidence cards with a synthesized anchor answer to select the final result. Experiments use a pool of six smaller specialist and generalist models across six reasoning benchmarks spanning science, mathematics, coding, finance, and medicine. ARGOT obtains the highest reported mean score and leads on four benchmarks. Ablations, disagreement-conditioned analysis, minority-answer rescue measurements, sensitivity tests, and qualitative success/failure cases explore routing and verification mechanisms.

Critical strengths
  1. Explicit separation of support and challenge, with protected minority edges, targets an important failure mode in majority-dominated collaboration. The framework requires no training and makes final selection more inspectable through trajectory evidence cards.

  2. The evaluation covers six benchmarks and provides a controlled comparison with GoA using the same heterogeneous model pool and dynamic agent selection. Disagreement-conditioned analysis and minority-answer rescue analysis are relevant to the stated motivation.

Critical weaknesses
  1. Existing ablations do not establish the necessity of optimal transport. For multiple-choice tasks, transport distributions depend only on selected options and self-reported confidence, making them a coarse proxy for claim-level evidence.

  2. The evaluation does not sufficiently separate evidence routing from the stronger final-verification stage. Removing verification causes the largest drop on HumanEval, but the replacement is majority voting and GoA lacks the same trajectory verifier.

  3. Computational trade-offs remain unclear because end-to-end token, model-call, and latency comparisons are missing. ARGOT involves selection, two directed refinement stages, anchor-answer synthesis, and final verification, while baselines retain different interaction pipelines. Matched inference-budget comparisons are required to determine whether accuracy gains reflect better routing or additional computation.

Questions for authors

See Weaknesses.

3 comments

Many thanks to Reviewer fYno for appreciating the significance of our methodological innovations and key mechanisms. To address your insightful suggestions, we have conducted further empirical analyses to clarify the specific impacts of our underlying modules and assess the overall system efficiency.

W1: Necessity of Optimal Transport

We agree that, for multiple-choice tasks, the transport distributions represent the selected option and self-reported confidence without directly encoding the full supporting rationale. Their role is to provide a routing signal. The ground cost incorporates relationships between options, so transporting probability mass between semantically similar options incurs a different cost from transporting it between dissimilar options. The reasoning text is examined during subsequent communication and verification.

To assess OT’s contribution, we replace it with answer disagreement, JS divergence, or embedding similarity while retaining agent selection, verification, and the communication configuration.

Relation MetricGPQAHumanEval
OT (Ours)39.9085.98
Answer disagreement34.8581.71
JS divergence28.7979.88
Embedding similarity35.8681.10

As shown, OT consistently outperforms all alternative metrics. Compared with the strongest alternative on each benchmark, OT improves GPQA by 4.04 percentage points and HumanEval by 4.27 percentage points.

Although the transport distributions use selected answers and confidence scores, the task-adaptive ground cost captures semantic relationships between answer representations. These results suggest that OT provides a more effective basis for constructing support and challenge relations than simpler disagreement or similarity measures, even without explicitly encoding full reasoning trajectories into the transport distributions.

W2: Separating Evidence Routing from Trajectory Verification

Thank you for highlighting this concern. We agree that comparing against majority voting and the original GoA does not fully isolate the contribution of evidence routing when the verification stage differs. To address this concern, we added two controlled comparisons: one applies the same trajectory verifier to GoA, and the other directly verifies independent initial responses without evidence routing. These experiments allow us to better distinguish the contributions of routing and verification.

VariantRoutingGPQA
GoA + Trajectory VerifierOriginal GoA routing35.35
ARGOT (Verifier Only)None26.77
ARGOT (Ours)OT-based routing39.90

With the same final verifier, ARGOT outperforms GoA routing by 4.55 percentage points, indicating that the improvement cannot be attributed solely to the verification stage.

Furthermore, applying the verifier directly to independent initial responses yields only 26.77%, compared with 39.90% for complete ARGOT. This 13.13-point gain demonstrates the importance of structured evidence routing before final verification.

These comparisons support the complementary contributions of relational communication and trajectory verification, rather than attributing the performance gains to either component alone.

W3: Computational Trade-offs and Inference Efficiency

Thank you for raising this concern. We conducted an additional end-to-end efficiency comparison on the GPQA. We report accuracy, total input/output tokens, inference time, and average LLM calls per question.

MethodAcc. (%)Input (K)Output (K)Time (s)Calls/Question
MoA33.841332.31532.886754.597.00
GoA (Mean)35.351687.21446.136094.7411.19
ARGOT (Ours)39.901674.92260.204172.0612.60

As shown, ARGOT achieves the highest accuracy while reducing output tokens by 41.7% and inference time by 31.5% compared with GoA. Although ARGOT requires more LLM calls, its lower generation volume and shorter runtime suggest more efficient use of individual interactions.

These results demonstrate a favorable accuracy-efficiency trade-off, indicating that ARGOT's performance improvements are not simply driven by increased token generation or longer inference time.

Dear authors, could you clarify why the newly added GoA + Trajectory Verifier variant achieves exactly the same GPQA accuracy (35.35%) as the original GoA reported in Table 1?

Clarification on the GPQA accuracy for GoA + Trajectory Verifier

We sincerely thank the reviewer fYno for the careful reading and for raising this question. The identical accuracy is a coincidence at the aggregate level rather than a reuse of the original GoA predictions. Both variants correctly answer 70 out of 198 GPQA questions, resulting in the same accuracy of 35.35%. Their per-question predictions nevertheless differ substantially, and the changes introduced by the trajectory verifier happen to balance out in the aggregate accuracy.  

The result indicates that the verifier substantially changes the decision outcomes but does not provide a net improvement when combined with the original GoA routing on GPQA. In contrast, its stronger performance under ARGOT's OT-based relational routing suggests that the benefit depends on the interaction between relational routing and final verification.

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

This paper addresses a failure mode of multi-agent LLM reasoning that the authors call dominant-answer collapse: when most agents converge on a wrong answer, answer-level voting or free-form debate suppresses a correct minority trajectory. The proposed framework, ARGOT, (i) selects k=3 agents from a pool of six 7B–8B models via model cards, (ii) maps each agent's initial response to a distribution over answer options or tokens and uses entropic optimal transport to compute agent-to-agent and agent-to-consensus costs, (iii) builds a small relational graph whose edges are typed as support or challenge, with one protected challenge edge from the most divergent confident agent to the consensus anchor, (iv) runs one round of relation-conditioned message passing, and (v) replaces voting with a Qwen2.5-7B verifier that reads per-agent evidence cards. Experiments on six benchmarks (MMLU-Pro, GPQA, AIME, HumanEval, FinQA, MedMCQA) against six baselines report the best average accuracy and first place on four benchmarks, supported by ablations, a minority-rescue analysis, threshold sensitivity, and case studies.

Critical strengths
  1. Clear and well-motivated problem framing. The paper isolates a concrete failure mode (a correct minority suppressed by a wrong majority), connects it to prior observations on conformity in debate, and designs every stage of the pipeline around it. The minority-rescue analysis (Fig. 3b) and the case studies in Appendix F test this mechanism directly rather than only reporting aggregate accuracy. This matters for my assessment because it means the paper asks a well-posed question that the community cares about, and the proposed OT-based support/challenge typing is, to my knowledge, a new way to attack it.

  2. Comparatively honest reporting. The main table uses the stronger GoA variant (GoA_Mean) rather than the weaker GoA_Max, the paper states that self-reported confidence is uncalibrated, and Appendix F.4 includes failure cases. These choices make the comparison more credible than it would otherwise be and suggest that the problems listed below are fixable in principle.

Critical weaknesses
  1. Internal inconsistencies in the reported numbers undermine the main empirical claim. Several reported numbers are inconsistent: MoA’s MedMCQA accuracy (54.08) is incompatible with the stated 1,000-example evaluation and duplicates its MMLU-Pro result; Figures 4 and 7 disagree with the tables; and AIME accuracies imply 45 problems rather than the stated 60. Moreover, results use a single seed without uncertainty estimates, and several headline gains amount to only a few correct answers. These issues make the claim of “best average and first on four of six benchmarks” difficult to verify.

  2. The gains cannot be attributed to the claimed contribution, and the core scoring is not reproducible from the text. Removing the final verifier lowers GPQA accuracy from 39.90 to 34.34, below Debate (37.88), while baselines do not receive the same verifier. The experiments therefore do not establish how much of the improvement comes from the proposed graph mechanism. Reproducibility is also limited by missing definitions of the peer relevance score, OT ground cost, key coefficients, and support/challenge assignment rule. Additionally, the threshold is inconsistently applied to node versus edge scores. Matched-verifier comparisons and a complete scoring specification are needed.

  3. The novelty beyond GoA plus a verifier is limited.

Questions for authors
  1. Please provide the correct MoA MedMCQA value and the raw counts (correct / total) behind every entry in Tables 1, 2 and 7.
  2. Please explain the discrepancies between Figures 4 and 7 and Tables 1 and 7 listed above. Were the MMLU-Pro curves produced from a different run than the tables?
  3. Please explicitly define the peer relevance score b̄_i, the OT ground cost D, the coefficients λ, β and λ_cov, and the support/challenge assignment rule. Is threshold t applied to node scores or edge scores? The text and algorithm appear inconsistent.
4 comments

Many thanks to Reviewer ZDJz for providing an insightful review and recognizing our contributions. Guided by your insightful suggestions, we have made corresponding revisions to further improve the clarity and accuracy of our manuscript.

W1 & Q1: MoA MedMCQA Accuracy and Raw Counts

The correct MoA accuracy on MedMCQA is 54.50% (545/1000). The reported 54.08% was mistakenly copied from its MMLU-Pro result. We apologize for this transcription error.

With this correction, MoA’s six-benchmark average is 42.72%, while ARGOT’s remains 44.81%, corresponding to an average improvement of 2.09 percentage points. ARGOT’s MedMCQA improvement over MoA is 0.70 percentage points. The corrected results preserve ARGOT’s highest observed average and first-place ranking on four benchmarks. Tables 2 and 7 are unaffected by this correction.

The correct/total counts underlying all entries in Tables 1, 2, and 7 are provided below. Percentages are calculated from these counts and rounded to two decimal places.

Table 1: Raw counts (correct/total)

MethodMMLU-ProGPQAAIMEHumanEvalFinQAMedMCQA
SC487/98068/1984/4591/16476/1147508/1000
Refine446/98067/1983/45139/164201/1147483/1000
ReConcile520/98064/1987/45108/164189/1147512/1000
Debate522/98075/1983/45137/164201/1147530/1000
MoA530/98067/1985/45140/164200/1147545/1000
GoA496/98070/1984/45144/164210/1147533/1000
ARGOT546/98079/1986/45141/164215/1147552/1000

Table 2: Raw counts (correct/total)

VariantGPQAHumanEval
ARGOT79/198141/164
w/o agent selection66/198135/164
w/o OT graph73/198138/164
w/o challenge edges72/198137/164
w/o evidence verification68/198114/164

Table 7: Raw counts (correct/total)

MethodMMLU-ProGPQAAIMEHumanEvalFinQAMedMCQA
GoA Mean496/98070/1984/45144/164210/1147533/1000
GoA Max430/98066/1983/45138/164149/1147481/1000
ARGOT546/98079/1986/45141/164215/1147552/1000

AIME evaluation scope. The AIME denominator is 45, comprising 15 problems from the AIME 2023 split used in our dataset and 30 problems from AIME 2024. All methods are evaluated on the same 45-problem list.

All accuracies are computed as correct/total and rounded to two decimal places.

W1 & Q2: Reconciling the figures and tables

The MMLU-Pro curves in Figures 4 and 7 use the same fixed random 50-question subset selected with seed 0 for computationally efficient hyperparameter analysis. All methods and parameter settings shown in these curves use this same subset. Tables 1 and 7 instead report results on the 980-question evaluation set.

The subset description was inadvertently omitted during manuscript refinement. We have now clarified in the captions and accompanying text of Figures 4 and 7 that the MMLU-Pro curves use a fixed random 50-question subset solely for hyperparameter-trend analysis.

W2 & Q3: Methodological Definitions and Graph Construction

Thank you for identifying these ambiguities. We agree that the original description was insufficiently precise for reproduction. The implemented scoring and graph-construction procedure is specified below.

1. Peer relevance bˉi\bar{b}_i. Each agent aja_j assigns normalized peer scores pj→ip_{j\rightarrow i} based on correctness, coherence, and relevance. The mean received score bib_i is normalized to obtain bˉi\bar{b}_i.

2. OT ground cost D\mathcal{D}. We construct a task-adaptive semantic ground-cost matrix to capture discrepancies between answer representations. For multiple-choice tasks, D\mathcal{D} uses normalized cosine distances between option-text representations, with a binary fallback when parsing fails. For open-ended tasks, normalized Jaccard distances are computed over task-specific feature tokens for mathematical, code, and textual reasoning. This allows OT to capture task-aware distributional disagreement beyond simple label mismatches.

3. Coefficients. λ>0\lambda>0 controls the exponential decay of consensus reliability with transport divergence, while β∈(0,1)\beta\in(0,1) bounds the contribution of confident disagreement.

λcov\lambda_{\mathrm{cov}} was a conceptual generalist-coverage term, not a tuned coefficient. We have removed it and now use MetaRank⁡+TopK⁡\operatorname{MetaRank}+\operatorname{TopK} for agent selection, with generalist coverage retained in the meta-agent ranking criteria.

4. Threshold and relation assignment. The threshold tt applies only to node scores bˉi\bar{b}_i, defining

St=ai∈S:bˉi>t.\mathcal{S}_t=\\{a_i\in\mathcal{S}:\bar{b}_i>t\\}.

Among active agents, answer agreement is typed as support and disagreement as challenge. The highest-diversity agent disagreeing with the consensus anchor is assigned a directed protected-challenge edge, ensuring that reliable minority evidence survives pruning. Edge scores remain relational metadata, while message weights are normalized from source-node scores.

W2 & W3: Separating routing from verification and clarifying the contribution beyond GoA

We acknowledge the substantial contribution of trajectory verification, particularly on HumanEval, as reflected in the original ablation results. To further isolate the benefit of relational communication under a shared verifier, we conducted additional controlled comparisons. Our additional mechanisms concern relation-conditioned communication and the preservation of a selected challenger before trajectory-level verification.

To examine the contribution of routing under a common final verifier, we evaluate the following controls:

VariantGPQA
GoA + the same trajectory verifier35.35
Independent responses + the same trajectory verifier26.77
ARGOT39.90

These variants use the same agent pool, evaluation split, verifier model, verifier prompt format, and decoding settings. For the verifier-only variant, the verifier receives the independent initial trajectories in the same candidate format but without graph communication or route-specific relation metadata.

ARGOT outperforms GoA with the same verifier by 4.55 percentage points and direct verification of independent responses by 13.13 points on GPQA. These observed differences indicate that the gain cannot be attributed solely to final answer selection: structured communication and refinement before verification also contribute substantially.

We further replace only the relation metric within the ARGOT pipeline:

Relation metricGPQAHumanEval
OT39.9085.98
Answer disagreement34.8581.71
JS divergence28.7979.88
Embedding similarity35.8681.10

For this comparison, the selected agents, verifier, communication budget, decoding settings, and downstream graph-processing procedure are held fixed; only the metric used to construct support and challenge relations is replaced.

OT exceeds the strongest tested alternative by 4.04 percentage points on GPQA and 4.27 points on HumanEval. The metric comparison isolates OT within the routing stage, whereas the common-verifier controls separate routing and communication from final answer selection. Together, these results support complementary contributions from relational evidence routing and trajectory verification.

Thanks for your reply. I’ve increased my rating from 1 to 2.

Thank you for your time in reading our response and for updating the rating. We remain at your disposal if you have any further questions or need additional clarifications.

Reviewer_AsvXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARGOT, a multi-agent reasoning framework that constructs support and challenge relations between agents using OR. It selects task-relevant agents, converts their initial answers into distributions, and uses transport costs to guide relation-aware communication while preserving a minority challenger. A trajectory-level verifier then selects the final answer. Experiments on six benchmarks show the highest average performance among the compared methods, with additional analyses of disagreement and minority-correct cases.

Critical strengths
  1. Preserving useful disagreement is a sensible objective for heterogeneous multi-agent systems. The paper goes beyond overall accuracy by examining cases with weak initial consensus and cases where the correct answer is initially in the minority. These analyses help connect the evaluation to the stated motivation.

  2. ARGOT performs competitively across several reasoning domains and achieves the best results on four of the six benchmarks. The component ablations also provide useful evidence that agent selection, communication structure, and final verification each affect performance.

Critical weaknesses
  1. The benefit of OT for evidence routing is not clearly established. For multiple-choice tasks, Eq. 2 represents each response using only the selected option and self-reported confidence, without directly encoding the supporting rationale. Although the ground cost incorporates option-level semantic similarity and later stages inspect the reasoning text, it remains unclear what OT adds beyond simpler answer- and confidence-based scores or alternative distances such as JS divergence and embedding similarity. The current ablation replaces the OT graph with a fully connected graph, changing the communication structure at the same time. A comparison that holds the rest of the pipeline fixed would more directly establish whether OT is necessary.

  2. The contribution beyond GoA and the final verifier needs to be better separated. ARGOT follows GoA's agent-selection procedure and retains a similar selection, graph communication, and aggregation pipeline. Its main additions are OT-based relations, protected challenge edges, and trajectory verification. However, the current comparisons do not establish how much the routing contributes once the verifier is held fixed. On HumanEval, removing verification reduces performance from 85.98% to 69.51%, whereas replacing the OT graph reduces it to 84.15%. A GoA baseline with the same verifier, and a baseline that directly verifies the same independent initial responses without communication, would be particularly informative.

Questions for authors
  1. With the selected agents, verifier, and communication budget held fixed, how does OT compare with answer disagreement, JS divergence, or embedding similarity for constructing the support and challenge relations?

  2. How do GoA with the same trajectory verifier and independent agent responses followed directly by that verifier perform? These comparisons would help separate the benefit of routing from that of final answer selection.

5 comments

We thank Reviewer AsvX for recognizing the value of preserving useful disagreement and our disagreement-focused analyses. Your comments identify two important attribution questions, which we address through additional controlled comparisons below.

W1 & Q1: Contribution of OT relative to alternative relation metrics

We agree that, for multiple-choice tasks, Eq. (2) represents the selected answer and self-reported confidence; it does not directly encode the supporting rationale. OT provides a routing signal over these distributions: its ground cost distinguishes moving probability mass between semantically similar options from moving it between dissimilar options. JS divergence on the same distributions does not incorporate this cross-option geometry. The reasoning text is subsequently examined during communication and verification.

We acknowledge that the original fully connected graph ablation did not fully isolate the contribution of the OT relation metric from changes in graph topology. We therefore compare OT with answer disagreement, JS divergence, and embedding similarity within the same ARGOT pipeline, holding the selected agents, verifier, communication budget, and remaining relation-construction procedure fixed.

Relation metricGPQAHumanEval
OT (ARGOT)39.9085.98
Answer disagreement34.8581.71
JS divergence28.7979.88
Embedding similarity35.8681.10

OT achieves the highest score on both benchmarks, exceeding the strongest alternative on each by 4.04 percentage points on GPQA and 4.27 points on HumanEval. These controlled comparisons support the added value of OT over the tested alternatives within ARGOT’s routing and verification pipeline. OT’s role is to guide evidence routing; the correctness of the underlying reasoning is assessed in subsequent stages.

W2 & Q2: Separating relational routing from final verification

We agree that ARGOT adopts GoA’s agent-selection procedure and that final verification contributes substantially, particularly on HumanEval. The original ablations alone do not establish the incremental benefit of routing under a shared verifier. We therefore evaluate the two requested controls:

  • GoA + trajectory verifier: retains GoA’s routing and replaces its final aggregation with the same trajectory verifier used by ARGOT.
  • Independent responses + trajectory verifier: applies that verifier directly to the independent initial responses, without inter-agent communication.
VariantRoutingGPQA
GoA + trajectory verifierOriginal GoA routing35.35
Independent responses + trajectory verifierNone26.77
ARGOTOT-based relational routing39.90

With the same final verifier, ARGOT improves over GoA by 4.55 percentage points and over direct verification of independent responses by 13.13 points on GPQA. The first comparison supports an additional benefit from ARGOT’s relational routing over GoA’s routing; the second supports the value of communication and refinement before final verification.

Together, these results support complementary contributions from routing and verification on GPQA. They are consistent with the substantial role of verification observed on HumanEval, while showing that a shared final verifier does not eliminate the performance differences between routing strategies.

Thank you for the additional results. It substantially address my concern about the contribution of OT and help separate routing from verification on GPQA. If available, could you also report the two shared-verifier comparisons on HumanEval? My original concern was motivated particularly by the large performance drop when verification was removed on that benchmark, so these results would help clarify the respective contributions of routing and verification.

Follow-up to W2 & Q2: Shared-verifier comparisons on HumanEval

We sincerely thank the reviewer AsvX for the positive feedback and the constructive follow-up request.

To address the concern regarding the performance drop on HumanEval when verification is removed, and to further clarify the respective contributions of routing and verification, we have evaluated the two requested shared-verifier controls:

VariantRoutingGPQA Acc. (%)HumanEval Pass@1 (%)
GoA + trajectory verifierOriginal GoA routing35.3578.05
Independent responses + trajectory verifierNone26.7778.66
ARGOTOT-based relational routing39.9085.98

These results clarify the mechanism behind the ablation drop and demonstrate that ARGOT's routing and verification are highly synergistic:

1. The role of verification and the ablation drop: The independent baseline (78.66%) confirms that verification establishes a critical performance floor on strict coding tasks. The substantial drop (to 69.51%, as reported in our original ablation study) without verification is an expected consequence of ARGOT's design: OT-based routing deliberately preserves informative disagreements and diverse minority paths to expand the search space. Without the trajectory verifier to rigorously filter this intentionally diverse candidate pool, the retained disagreements act as noise.

2. The structural mismatch of conventional routing: Conversely, while conventional GoA routing improves the low independent baseline on GPQA, it creates a structural mismatch on strict coding tasks, slightly degrading the strong independent baseline (78.05% vs. 78.66%). GoA's native routing mechanism tends to drive early consensus and homogenize responses, whereas our verifier requires structurally diverse reasoning graphs to be maximally effective. By actively maintaining high-quality diversity, ARGOT’s OT-routing successfully breaks this bottleneck, outperforming independent verification by 7.32 points and the GoA routing variant by 7.93 points.

Together, these comparisons demonstrate that while verification is necessary to ensure logical correctness, OT-based routing is strictly essential to organize productive communication and provide the high-quality, diverse candidates required to break the performance bottleneck.

We hope these additional results and the mechanism analysis fully clarify the complementary roles of both components.

Comments on the paper

1 comment

Summary of Rebuttal

We sincerely thank the Area Chair and all reviewers for their time and constructive engagement throughout the discussion period. As the discussion concludes, we briefly summarize the key clarifications and additions made during the rebuttal:

  • Disentangling OT Routing from Verification: To separate the contribution of OT-based routing from final verification, we added shared-verifier controls (GoA + verifier and Independent + verifier) on GPQA and HumanEval. Under the same verifier, ARGOT outperforms both GoA + verifier and Independent + verifier on these benchmarks, illustrating the complementary roles of relational routing and trajectory verification.
  • Relation Metric Ablations: We compared OT against alternative relation metrics (answer disagreement, JS divergence, and embedding similarity) within the same pipeline. The results show that OT achieves better performance across both benchmarks, supporting its usefulness for constructing support and challenge relations.
  • Inference Efficiency: We added an overall efficiency comparison on GPQA, showing that ARGOT achieves higher accuracy with fewer output tokens and shorter inference time than the compared baselines, indicating a favorable balance between accuracy and efficiency.
  • Data Consistency & Clarity: We corrected the MedMCQA transcription typo, provided the raw counts for the main and ablation tables, clarified the subset used for the hyperparameter curves, and supplied explicit definitions for the scoring and graph-construction procedures.

Concluding Remarks: We hope these additional experiments and clarifications have addressed the main technical, empirical, and clarity concerns raised in the initial reviews, and we are grateful for the positive reassessments and score increases following the discussion. We hope ARGOT offers a useful contribution to multi-agent reasoning, and we thank the AC and reviewers again for helping improve this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.