acceptodds
Withdrawn from ICLR 2027

Learning Informative and Efficient Multi-Agent Communication Topologies via Fine-Grained Information Gain Rewards

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The communication topology in large language model (LLM)-based multi-agent systems governs how information flows among agents, critically shaping both task performance and communication efficiency. However, existing topology learning methods typically rely on graph-level outcome rewards, which provide limited feedback on individual edge contributions. Consequently, even successful topologies may retain redundant or harmful interactions, introducing unnecessary communication overhead and potentially hindering collaboration. To address this limitation, we propose EdgeIG, an information-theoretically guided framework for fine-grained reward modeling in adaptive communication topology learning. Specifically, EdgeIG guides topology generation policy learning based on the information gain from each transmitted message, which is measured by how that message changes the model’s support for the ground-truth answer. The resulting edge-level rewards provide dense supervision, penalizing harmful interactions even within successful topologies. We further combine these rewards with a graph-level advantage to jointly improve task performance and communication efficiency. Extensive experiments on six benchmarks spanning general reasoning, mathematical reasoning, and code generation demonstrate that EdgeIG improves task performance and communication efficiency while maintaining strong robustness under prompt attacks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Model Agents, Multi-Agent Systems, Topology Optimization, Fine-Grained Credit Assignment

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_7efuOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper argues that existing methods evaluate whether a communication topology is good based on the final outcome of the entire communication graph, which may overlook the importance of individual communication edges. Therefore, the paper trains a framework that removes communication edges one by one from a counterfactual perspective and evaluates the importance of each edge based on the change in the outcome before and after removal.

Critical strengths
  1. The storyline is clear and the problem is well defined. Evaluating a graph solely based on the final outcome can indeed overlook the importance of individual communication edges.

  2. The method is intuitive and consistent with how humans might approach the problem. The counterfactual idea is easy to think of and straightforward to understand.

  3. The experiments cover a broad range of settings.

Critical weaknesses
  1. Removing individually redundant communication edges may reduce system stability or robustness, even if task performance is unaffected.

  2. The first-order counterfactual analysis cannot capture dependencies among multiple edges and therefore cannot determine whether individually removable edges are jointly removable.

  3. The edge-level credit assignment may be overly fine-grained, as the importance of communication can arise from interactions among groups of edges rather than individual edges.

  4. Because LLM outputs are stochastic, the observed performance change after removing an edge may partly reflect sampling variance rather than the causal effect of that edge.

Questions for authors

Thank you for your paper. I have a few questions about it.

  1. In information theory, redundant information may help maintain system stability. If redundant communication is removed simply based on whether it improves the final outcome, could this affect the stability or robustness of the system?

  2. Suppose you find that removing A, B, or C individually does not affect the result. However, it could be that at least two of A, B, and C are needed, or that just one of them is enough. How do you distinguish between these cases? Do you need to enumerate different combinations? If we move from first-order counterfactuals to higher-order counterfactuals, where multiple edges are removed at the same time, the computational cost will increase significantly. Is the improvement worth such a large cost? The 97.26% result in Figure 1 can only show that these edges are individually removable, but not that they are jointly removable.

  3. The paper tries to solve the problem that graph-level credit assignment is too coarse-grained, but does it go too far in the other direction? Sometimes what matters is not whether A or B is important individually, but whether A and B are important together.

  4. Since LLM outputs themselves have variance, is it possible that the change observed after removing an edge in the counterfactual setting is simply caused by resampling variation rather than the effect of removing that edge?

Reviewer_cCaDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This submission studied how to refine cooperation topology that can be adaptive to the given task. They design a new reward function that consider incremental edge gains, penalty on harmful interactions and with other general items to guide the policy leanring.

Critical strengths

None

(I carefully read this submission. Although nice presentation, I did not find remarkable contributions that can supass the standard line of this confernece. The core contribution of this submission is the authors design a new reward function and integrate it to the standard agent policy framework. In the past two years, there are enormers works on communication topology of agents. I think such contribution is not novel and not strong enough. )

Critical weaknesses

Please see my comment in the strength comment.

The main concern is that the innovation of this submission is very limited. Beyond that, I have not found further more serious weaknesses, but there are still some minor weaknesses that may make this submission more solid, including more extensive experiments, theory analysis on why it works.

Questions for authors

None

Reviewer_PFM2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EdgeIG assigns rewards to individual communication edges by measuring how (1) keeping or (2) removing a message changes support or the results for the correct answer. In my personal opinion, it addresses a credit-assignment problem, but i think the overall novelity is marginal and the evaluation and interpretation of that reward need strengthening.

Critical strengths
  • A successful final answer does not establish that every communication edge was useful. The preliminary edge-deletion experiment provides a concrete motivation for more granular supervision.
  • Handing out feedback for specific messages along communication paths makes a lot of sense conceptually. The GraphIG comparison in Table 2 helps test whether allocating feedback at the edge level is beneficial.
Critical weaknesses
  • A change in the decision model’s likelihood of a reference answer does not necessarily measure an edge’s contribution to final system correctness. Resampling the receiver after removing a message also introduces generation variability. Downstream reward summation adds another approximation. I would want evidence that edge scores predict actual end-to-end gains or losses, together with an analysis of reward noise.
  • The main experiments use one backbone, Qwen3-8B. MMLU contains only 153 test examples, and several improvements over the strongest per-dataset baseline are around one percentage point or less. I think more experimentation would be really helpful
  • In my personal opinion, I think this work is really close to https://arxiv.org/pdf/2510.14967, and I assume this work built on top of this work?
Questions for authors

see weakness

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

Related papers

Discussion (0)

Sign in to comment.