acceptodds
Under review as a conference paper at ICLR 2027

BiGate: Bidirectional Verification-Guided Selective Communication for Multi-Agent Reasoning

Abstract

Multi-agent reasoning can exploit complementary capabilities across heterogeneous language models, but indiscriminate full-reasoning exchange increases inference cost and exposes revision contexts to unnecessary or erroneous peer information. We propose BiGate, a bidirectional verification-guided selective communication mechanism. For each agent pair, BiGate forms a response-local ternary verification relation from the two directional judgments. The Sender Need Gate uses the sender’s judgment of the receiver to decide whether communication is needed and whether feedback is supplementary or corrective, while the Receiver Scope Gate uses the reverse judgment to control whether the sender’s full reasoning enters the receiver’s revision context. Their joint mapping yields dynamic directed typed communication, while two-stage context isolation prevents verification-time peer reasoning from carrying into revision except through gated messages. Across five reasoning benchmarks and three heterogeneous base models, BiGate achieves 73.90% macro-average accuracy, 2.30 percentage points above the strongest collaborative baseline, and uses the fewest model calls and Response Rounds among the evaluated multi-agent methods. Ablations and behavioral analyses support the gates’ distinct roles and show task-varying communication with restricted full-reasoning admission under persistent error.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi-agent reasoning, large language models, selective communication, bidirectional verification, adaptive communication, heterogeneous LLMs, communication efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GfCzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BiGate proposes selective communication for heterogeneous multi-agent reasoning. Each agent pair receives directional ternary judgments. The sender's judgment determines whether feedback is needed, while the receiver's judgment determines whether the sender's full reasoning can accompany that feedback. Separate verification and revision contexts enforce this distinction. Experiments use three models and five benchmarks, with 200 sampled examples per benchmark. The paper reports 73.90% macro-average accuracy versus 71.60% for Peer Review, alongside fewer reported model calls and response rounds. Gate ablations and a persistent-error analysis examine how the mechanism changes communication and restricts admission of erroneous full reasoning into subsequent answer revision.

Critical strengths
  1. Clear separation of two communication decisions. The sender controls whether feedback is needed, while the receiver controls whether full reasoning is admitted. Separate revision contexts make this distinction operational and preserve potentially useful local criticism from untrusted solutions.

  2. Relevant comparisons and targeted ablations. The evaluation uses shared models and examples across five benchmarks. Table 4 compares both gates individually and jointly, with BiGate reporting higher accuracy and lower token usage than either single-gate configuration.

Critical weaknesses
  1. Call accounting appears inconsistent with the stated protocol. Each revision stage requires three verification calls and at least one revision call. Thus, 2.48 response rounds imply at least
3+4(2.48−1)=8.923+4(2.48-1)=8.92

calls, exceeding the reported 8.80. Caching, skipped verification, or a different round definition might explain this, but none is specified. This affects the central efficiency claim.

  1. The incremental accuracy benefit needs stronger support. Peer Review and FullConnect have identical aggregate accuracy and cost metrics, yet their implementation relationship is unclear. Moreover, the gains over Peer Review and Sender-Only---2.30 and 1.50 percentage points---lack paired uncertainty estimates; the two reported seeds concern sample selection rather than repeated generation.
Questions for authors

1.Can you reconcile the reported calls and response rounds by providing separate initial-generation, verification, and revision counts? Please clarify whether termination checks count as response rounds and whether verification is cached or skipped.

2.Are Peer Review and FullConnect the same implementation? Please clarify their relationship and provide paired uncertainty estimates for BiGate versus Peer Review and Sender-Only.

Reviewer_AE9GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a communication protocol for selectively using peer information in multi-agent reasoning. Bidirectional verification determines whether an agent needs supplementary or corrective feedback and whether the sender’s full reasoning should be included. This allows agents to consider potentially useful local feedback even when they do not endorse the sender’s overall solution. Separate verification and revision contexts control which information enters the revision process. The method is evaluated across reasoning benchmarks with heterogeneous models, and further conducted gate ablations and analyses of communication behavior.

Critical strengths
  • BiGate presents a well-engineered framework for selective communication and cost control in multi-agent reasoning.

  • An interesting strength is that BiGate allows agents to consider a peer’s local feedback even when they do not endorse that peer’s overall solution, since a specific criticism may still be useful even when the sender’s own answer is unreliable.

Critical weaknesses

The empirical validation is limited. An ablation of the Supplement/Correction mechanism would help establish the value of distinguishing these feedback types. Evaluation on larger frontier models, and more analysis of how communication patterns change as the number of rounds increases, would strengthen the evidence for the method’s generalization and behavior.

My concerns at this stage mainly relate to the empirical validation. I would be willing to raise my score if the authors address these concerns.

Questions for authors

Could the authors provide some representative successful and unsuccessful trajectories? I am curious about the internal workings of BiGate.

Reviewer_urm2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a message-passing schema for better communication and convergence across heterogeneous LLMs. The schema is based on bidirectional flow of messages between agents where a message is inspected by agents to get verifies or to find issues. Authors study the performance on the schema on across five reasoning benchmarks and three heterogeneous base models showing better average performance compared to baselines.

Critical strengths
  • Paper is easy-to-follow and the approach makes sense.
Critical weaknesses
  • Experiments are on 2 seed which may not give the full picture when two seeds are tied. also confidence intervals are not reported.
  • The models are small LMs and it is not clear how much the approach is helpful on larger / frontiers.
  • The results are not conclusive on benchmarks and the approach is better than baselines in 1 out of 5 benchmarks.
  • Also given that different model families will make different mistakes, that would be helpful to see when the message evaluation itself is incorrect and how it impact the whole dynamic.
Questions for authors

see the weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.