- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes RCGP, a query-conditioned graph policy that jointly determines which LLM agents participate, their roles, and their communication links. Starting from a sampled initial graph, it encodes separate receiving and sending role contexts to guide parallel node and edge predictions. The policy is initialized using synthetic collaboration graphs and trained through execution-guided reinforcement learning, rewarding correct answers and sparse communication. Experiments on six benchmarks covering reasoning, mathematics, and code generation report improved task performance, with additional analyses showing favorable token efficiency.
Critical strengths
The paper is clearly written and well organized. The experiments closely match the main claims and support the proposed design.
Critical weaknesses
A key logical link is insufficiently justified: why role contexts from an initial graph unrelated to the query should help generate the final collaboration graph.
Questions for authors
The paper is clearly presented, and the experimental results look solid overall. My main concerns currently center on the initial graph. If the authors can satisfactorily address the following questions, I would raise my rating.
-
Why is an initial graph needed? Is it required by the execution framework, or is it introduced for a particular modeling purpose?
-
Since the initial graph is not conditioned on the query, why should its neighborhood role statistics help predict the collaboration structure needed for that query? What advantages does random graph initialization offer over a fixed graph or learnable slot representations?
-
How sensitive are the results to the sampled initial graph? For the same query, could different initial graphs lead to substantial variations in team composition, answer accuracy, or execution cost?
-
The initial connections have not been validated for the current task, the roles can be reassigned, and the final neighborhoods may change entirely. Why, then, are the initial neighborhood relations useful for describing the final team's functional context? The role-context ablation supports the module's usefulness, but the connection between relations in a random candidate graph and meaningful functional context for the task remains insufficiently explained.
6 comments
Response to Q4:
As explained in our response to Q2, different initial graphs provide different starting relationships. During training, execution feedback teaches the policy how to use these candidate role contexts. At inference, the trained policy combines them with the query to predict the final graph in one pass. The initial roles and connections need not be preserved: their descriptors guide graph prediction rather than directly describe the final team’s functional dependencies. We will revise the wording to make this distinction explicit.
Dear Reviewer,
We hope this message finds you well. We have provided detailed responses to each of the concerns raised in your review, and we believe they address all of the points you mentioned.
With less than two days remaining in the discussion period, we would like to kindly ask whether you have any further questions or concerns. If so, we would be more than happy to address them promptly.
If our responses have resolved your concerns, we would sincerely appreciate it if you could kindly consider raising your score.
Thank you very much for your time and effort in reviewing our paper.
We sincerely thank the reviewer for these comments, which helped us improve the clarity of the paper. The clarifications above are now included in the revised version.
Response to Q1:
The initial graph serves a modeling purpose; only the final active graph is executed. It provides role-labeled neighborhoods and a shared random input for node–edge prediction. Without an initial graph, the factorized policy in Eq.~9 would condition only on the query and deterministic slot features and sample raw node and edge actions independently. We instead condition all predictions on the same sampled initial graph. Marginalizing over initial graphs yields a mixture of factorized policies that can capture action dependencies while retaining one-pass parallel decoding. Initial graphs are also used in prior work, such as the anchor topology in G-Designer (Zhang et al., 2025d), the communication graph that AgentPrune (Zhang et al., 2025c) prunes, and the noise graph that GTD (Jiang et al., 2026) denoises.
Response to Q2:
We do not assume that the initial graph already matches the query. Its statistics describe candidate role relationships, which we combine with the query to predict the final collaboration graph. For example, for the electricity-cost query in Fig.~2, one initial graph may connect the Math Analyst to the Economist, while another connects it to the Critic. These provide different starting relationships; the policy learns which roles and connections to retain or change according to the query and task-level execution feedback. The initial connections therefore need not be correct or preserved.
To assess this initialization choice, we compare random initialization with fixed Chain, Mesh, and Edgeless initializations. As shown in the table below, the average score decreases by 1.56% with Chain, 1.59% with Mesh, and 1.69% with Edgeless, relative to sampled initialization. Among fixed structures, Chain performs best among fixed structures on MMLU, whereas Mesh performs best on AQuA and HumanEval, indicating that the preferred fixed structure varies across datasets. Edgeless initialization sets the neighborhood statistics to zero but still permits edge prediction; it yields the lowest average score. Sampling instead supplies shared randomness that can capture dependencies across raw node and edge actions after marginalization while retaining parallel prediction. Overall, these results suggest that the policy benefits from varied starting relationships rather than always receiving the same initial connectivity pattern, providing an effective starting point for query-conditioned graph-policy learning.
Response to Q3:
We evaluate 133 MMLU queries with ten initializations each, totaling 1,330 executions. As shown in table below, the correctness outcome remains unchanged across all ten initializations for 77.44% of queries. We measure changes in team composition and connectivity using role-aware graph edit distance, which accounts for role changes and node/edge additions or deletions under optimal node alignment. The mean normalized distance is 0.325. The median within-query token coefficient of variation is 35.36%, counting all agent calls and the final aggregation call. These results show that different initializations can produce different teams and execution costs, while most queries retain the same correctness outcome.
Fig. 6 in Appendix E.8 of our submission shows two initializations of the same query whose sampled initial graphs differ substantially in both roles and connectivity: the first contains four active agents (Doctor, Critic, Psychologist, and Historian), whereas the second contains only a Doctor and a Critic. RCGP nevertheless generates closely related final graphs: in both cases, a Lawyer and a Historian pass messages to a third agent, and the two graphs differ only in that agent's role (Historian vs. Psychologist). Notably, neither final graph retains the Doctor or the Critic, and the Lawyer does not appear in either initial graph. Both executions answer correctly, consuming 1,128 and 1,117 tokens, respectively. Across all ten initializations of this query, all executions answer correctly. This example illustrates that RCGP maps diverse initial graphs to similar collaboration structures with comparable execution costs.