acceptodds
Withdrawn from ICLR 2027

Runtime-Evolving Communication Topology Designer for LLM-Based Multi-Agent Collaboration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model (LLM)-based multi-agent systems (MAS) have demonstrated remarkable collective intelligence, largely enabled by carefully designed inter-agent communication topologies. However, existing methods typically determine communication topologies before execution, preventing them from adapting information flow to runtime messages and evolving reasoning states. This can result in both redundant communication and missed task-critical interactions. To address this limitation, we propose RE-Designer, a Runtime-Evolving Communication Topology Designer that formulates topology construction as a sequential decision-making process and optimizes it through reinforcement learning (RL). Specifically, RE-Designer incrementally constructs a directed acyclic graph by dynamically adding agents and communication edges based on the partially constructed graph and the runtime messages generated by preceding agents. Since optimizing these sequential decisions with sparse task-level rewards is challenging, we introduce an information-bottleneck-inspired process reward that promotes task-relevant communication while penalizing redundant information propagation. To further reduce the exploration space, we develop a three-stage training strategy that first learns structural and message-aware priors through supervised learning and then refines runtime topology decisions through online RL. Extensive experiments on multiple reasoning benchmarks demonstrate that RE-Designer consistently achieves superior task performance, with favorable accuracy–communication trade-offs.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Multi-Agent Systems, LLM-Based Agent

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_dMzuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RE-Designer builds a communication graph while agents are running, rather than fixing the whole graph beforehand. It selects agent roles, chooses which previous messages to pass on, and decides when to stop. Training combines supervised examples with RL and a semantic-similarity-based feedback signal. Experiments cover six reasoning and code-generation benchmarks.

Critical strengths
  • The main idea is practically useful. Letting actual intermediate outputs influence later communication is more flexible than deciding every connection before seeing any results. The method also considers when to stop, rather than only which edges to add.

  • The message-conditioning ablation directly tests the central idea. Starting from the same checkpoint, the authors compare policies with and without access to runtime messages. The additional improvement supports the claim that these messages help the controller make better decisions.

Critical weaknesses
  • The distinction from the closest related method is not fully established. Section 2 describes AnyMAC as selecting the next agent and relevant previous outputs during execution, but it is absent from the main comparison. It remains unclear how much additional capability or benefit comes from explicitly representing this process as graph generation.

  • The information-based feedback has a weak justification. I understand that it is only inspired by the information bottleneck. Still, similarity to the final answer does not necessarily mean that a message helped produce that answer, and similarity to the question does not necessarily mean redundancy. The ablation supports its usefulness, but not the claimed interpretation of what it measures.

Questions for authors
  • What does the explicit graph representation add beyond AnyMAC’s next-agent and context selection? Is there a comparison using the same backbone and a comparable token budget?

  • Do messages receiving higher information weights actually contribute more to final correctness? For example, does removing them hurt performance more than removing low-weight messages?

Reviewer_9a5zOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents RE-Designer, which interleaves agent execution with incremental construction of a directed acyclic communication graph. Runtime messages guide edge selection and stopping decisions. Its three-stage training pipeline combines supervised structural learning, message adaptation, and online reinforcement learning with an information-bottleneck-inspired process reward. Experiments using Qwen3-8B on six benchmarks evaluate task performance and communication efficiency.

Critical strengths

This paper studies adaptive multi-agent communication topology design during execution, which adapts information flow to runtime messages and evolving reasoning states. It formulates communication topology construction as a runtime sequential decision process, providing a clear framework for interleaving agent execution with role selection, edge generation, and stopping decisions.

Critical weaknesses
  1. Insufficient Novelty Justification. Recent work on execution-adaptive multi-agent topology design is missing in related work, such as MAS² [1] and EvoMAS [2], both of which incorporate execution-derived information into system adaptation. Their omission leaves RE-Designer’s claimed novelty insufficiently distinguished from existing work.

[1] MAS²: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems, ICLR, 2026.
[2] EvoMAS: Evolutionary Generation of Multi-Agent Systems, ICML, 2026.

  1. Potential Bias in the Reward Proxy. The process reward favors similarity to the system’s own aggregated response and penalizes similarity to the query, without directly validating these proxies as measures of intermediate contributions. For example, merely repeating the final conclusion may yield high similarity, while a crucial intermediate derivation may be less similar.

  2. Insufficient Comparison. Lack of static baselines in Table 1 to clarify the additional benefit of runtime topology adaptation, since the training data use chain, layered, mesh, and fully connected topology. Furthermore, all experiments use Qwen3-8B, leaving the effectiveness across different model and scales unverified.

  3. Missing Sensitivity Analysis. Lack of sensitivity analysis for β and η, which control the query-similarity penalty and information-weighting strength. The robustness of the reported gains to these choices remains unverified.

Questions for authors

Please see weaknesses above.

Reviewer_P15aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RE-Designer, a runtime-evolving communication topology designer for LLM-based multi-agent collaboration. It formulates topology construction as a sequential decision-making process using a 22.94M lightweight policy to incrementally construct a directed acyclic graph based on runtime messages. Through a three-stage training strategy and an information-bottleneck-inspired process reward, RE-Designer enables runtime-adaptive information flow, consistently achieving superior task performance with favorable accuracy–communication trade-offs.

Critical strengths
  • The framework adapts subsequent communication decisions to runtime messages, with controlled ablations supporting the benefit of message conditioning.
  • The reported results show a favorable accuracy–communication trade-off. On MMLU, RE-Designer achieves higher accuracy than OFA-MAS while using approximately 30% fewer prompt tokens.
Critical weaknesses
  • The supervised trajectories are drawn from a predefined template set, with trajectory weights based on structural compactness. The robustness of the learned policy to changes in template coverage and alternative weightings of successful trajectories remains insufficiently characterized.
  • The topology policy relies on a frozen, general-purpose sentence encoder. Its sensitivity to encoder choice and its reliability on long messages, specialized terminology, and subtle logical contradictions remain unclear.
Questions for authors
  • How sensitive is the topology policy to the choice of frozen sentence encoder, particularly when messages contain subtle logical errors?

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

Related papers

Discussion (0)

Sign in to comment.