acceptodds
Under review as a conference paper at ICLR 2027

GAMMA-LLM: Game-Theoretic Adaptive Multi-Agent Message Allocation for LLM Reasoning

Abstract

Multi-agent large language model systems can improve reasoning by combining independent hypotheses and intermediate feedback, but additional communication is not uniformly beneficial: redundant or poorly targeted interactions increase inference-time token usage and can propagate incorrect rationales. We propose GAMMA, a game-theoretic framework for adaptive multi-agent message allocation that estimates which agents should receive greater influence and which communication edges have positive marginal value under a specified communication objective. GAMMA models agents as players in a graph-restricted cooperative game, constructs offline supervision from node-level Myerson contributions and cost-aware conditional edge values, and amortizes these quantities into a lightweight graph policy for sparse communication and contribution-weighted aggregation. Across nine reasoning benchmarks, GAMMA achieves higher macro-averaged accuracy than the evaluated adaptive multi-agent baselines while using fewer online inference tokens. Ablation, calibration, and robustness analyses further show improved prediction of the game-derived targets, lower error propagation, and fewer wrong-consensus failures.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Multi-Agent Systems, Credit Assignment, Adaptive Communication Graphs, Cooperative Game Theory, Sparse Communication

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_F6xZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adaptive communication allocation in LLM-based multi-agent reasoning systems. The authors observe that increasing interaction among agents does not necessarily improve reasoning quality: redundant messages increase online token consumption, while poorly targeted communication can propagate flawed reasoning and reinforce incorrect consensus. To address these issues, the paper proposes GAMMA, which decomposes communication decisions into two components: estimating each agent’s contribution to the overall reasoning outcome, and evaluating the cost-adjusted marginal benefit of a communication edge in a given graph. In the offline phase, the authors perform multi-agent rollouts over different communication graphs and construct supervision targets using graph-restricted Myerson contributions and cost-aware conditional edge values. These signals are then distilled into a lightweight GNN. At inference time, the GNN predicts node and edge values from independent agent traces, selects sparse communication edges, and aggregates the updated answers using contribution-based weights.

Critical strengths
  1. Clear motivation and a meaningful research problem. The paper addresses two practical challenges, communication redundancy and error propagation.

  2. A well-structured game-theoretic supervision design. Modeling graph-restricted Myerson node contributions separately from conditional marginal communication gains helps distinguish the value of an agent from the incremental value of its messages.

  3. Broad benchmark coverage and informative core ablations. The paper compares GAMMA with single-model reasoning methods, static multi-agent protocols, and several adaptive topology methods.

Critical weaknesses
  1. A potential approximation gap between graph-conditioned edge-value targets and one-shot online edge selection. According to Eq. (9), the true value ΔΦ(e | G) of an edge depends on the current communication graph G. However, the online policy predicts candidate-edge values from independent traces and selects E* in a single pass, without re-estimating conditional gains as edges are added. Appendix H explains the use of conditional-mean estimates learned over the offline graph distribution, but these estimates need not equal the actual marginal gains in the final selected graph. If edges convey redundant or complementary information, positive individual predictions do not guarantee positive joint gains. The authors should quantify this approximation error and provide a targeted comparison with an iterative or graph-aware edge-selection strategy.

  2. Insufficient reproducibility details for constructing offline counterfactual supervision targets. A more detailed algorithm and verifiable computational-cost statistics would improve reproducibility.

  3. End-to-end costs are not quantified in the efficiency analysis. The main experiments report online inference tokens only.

  4. Statistical reliability and generalization across base models are not sufficiently demonstrated.

  5. The distinction from recent fine-grained credit-assignment methods could be clearer. For example, Graph-GRPO (Findings of ACL 2026).

Questions for authors

Does the graph-conditional approximation affect joint edge selection? When candidate-edge values are interdependent, how large is the gap between conditional-mean predictions over the offline graph distribution and the true marginal gains in the final selected graph?

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

This paper focuses on multi-agent message allocation; redundant or poorly targeted interactions increase inference-time token usage and can propagate incorrect rationales. They propose GAMMA, a game-theoretic framework for adaptive multi-agent message allocation that estimates which agents should receive greater influence and which communication edges have positive marginal value under a specified communication objective. Experiments show that GAMMA achieves higher macro-averaged accuracy than the evaluated adaptive multi-agent baselines while using fewer online inference tokens.

Critical strengths
  • This paper focuses on an important problem: reducing unnecessary communication and error propagation in multi-agent reasoning.
  • This paper performs comprehensive experiments to demonstrate effectiveness, including nine datasets across different domains and comprehensive ablation studies.
Critical weaknesses
  • Evaluation is limited to Qwen3.5-9B. It remains unclear whether the gains hold across different model families and sizes.
  • Edges with positive average gains may not remain useful when selected together. Since the policy learns average edge gains across different graphs, please check whether the selected edges still help in the final graph.
  • Token accounting is unclear. In Table 7, GAMMA uses fewer tokens than the majority-vote baseline despite still generating all six initial responses. Please clarify which generation and communication costs are included.
Questions for authors

None

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

GAMMA proposes an adaptive message allocation method for multi-agent LLM reasoning, aiming to determine which agents should receive greater answer weights and which communications are worth conducting. The authors use game-theoretic methods to estimate the contributions of agents and the benefits of communication after deducting costs during the offline phase, and train a lightweight graph neural network to select a small number of communications during inference, followed by a weighted aggregation of the agents’ answers. The experiments mainly use Qwen-3.5-9B and six agents with different roles, covering nine benchmarks across mathematics, knowledge, coding, question answering, and long-context reasoning. The experimental results show that, on the tested tasks, the method achieves higher accuracy than the evaluated adaptive multi-agent baselines while using fewer online inference tokens than these baselines.

Critical strengths

S1 When multiple agents collaboratively answer questions, it is necessary to determine whom to trust more and which communications help obtain correct answers. This paper separately learns the contributions of agents and the benefits of communication, and selects communications and aggregates answers accordingly. This design can improve issues such as information redundancy and error propagation in multi-agent communication, making it worthy of investigation.

S2 The experiments cover 9 benchmarks and compare the method with sixteen baselines. The authors also examine the roles of individual design components through ablation studies, test performance when one agent is replaced by a weaker model or adversarial agents are introduced, and report failure cases. These analyses help to understand the effectiveness and limitations of the method.

Critical weaknesses

W1 Before the method can be used, it is necessary to generate training labels and train the graph policy. However, although the paper presents the overall workflow, it does not clearly explain how many additional model calls are required for label generation, which computational results can be reused, or the total token consumption and elapsed time.

W2 The paper states that the training and evaluation data are separated, but it does not clearly specify the sources, sample sizes, and splitting methods of the training data for each task. In addition, the results do not report variations across repeated runs, making it difficult to determine whether the improvements are stable. Although the experiments cover a relatively wide variety of tasks, the main experiments use only Qwen-3.5-9B with different role prompts. Validation using other base models, as well as directly applying the trained policy to different tasks, is lacking.

Questions for authors
  1. What specific model calls are required to generate the training labels offline? Please report the total number of calls, token consumption, and time cost, and compare these with the preparation costs of AGP and G-Designer.

  2. Where does the training data used for the policy come from for each task, how much data is used, and how is it separated from the test data?

  3. Can the trained policy be directly applied when the task, base model, or agent roles change, or is it necessary to regenerate the labels and retrain the policy?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.