acceptodds
Under review as a conference paper at ICLR 2027

Learning Role-Context Graph Policies for Adaptive Multi-Agent Collaboration

Abstract

Effective LLM-based multi-agent collaboration requires jointly deciding which agents should participate and how they should communicate. Existing methods often optimize communication over a prescribed team, potentially leading to redundant or insufficient participation for individual queries, or adapt participation through selection from available structures or sequential graph construction, limiting structural flexibility or decoding parallelism. Moreover, limited modeling of local directional role relationships can obscure the distinct effects of information reception and dissemination. To address these limitations, we propose RCGP, a query-conditioned Role-Context Graph Policy for adaptive multi-agent collaboration. To reflect their structural coupling, we formulate participation, functional roles, and directed communication within a unified graph-policy action space, allowing collaboration scale and topology to adapt jointly through parallel node–edge decisions. We further design directional role-context encodings to characterize local role diversity and neighborhood support separately in receiving and sending contexts, enriching graph-policy prediction beyond role identity and connectivity while preserving the functional asymmetry between information reception and contribution dissemination. We directly optimize the graph policy via execution-guided reinforcement learning (RL) with a structure-aware initialization. Together, the graph policy provides RL with a structured and relational collaboration space to optimize, while RL grounds the graph policy with task-aware execution utility, calibrating collaboration toward query-specific effectiveness. Extensive experiments on six benchmarks demonstrate that RCGP outperforms state-of-the-art baselines, while additional analyses show favorable cost–performance trade-offs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multi-Agent Systems, Graph Learning

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_4PhSOpenReview ↗
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.

  1. Why is an initial graph needed? Is it required by the execution framework, or is it introduced for a particular modeling purpose?

  2. 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?

  3. 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?

  4. 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

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.

InitializationMMLUHumanEvalAQuA
Random (ours)84.2192.1185.29
Chain83.4691.2382.84
Mesh81.9591.6783.82
Edgeless83.0890.7983.33

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.

MetricResult
Queries with unchanged correctness (%)77.44
Normalized role-aware graph edit distance0.325
Within-query token CV: median (%)35.36

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.

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.

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

This paper proposes RCGP, a query-conditioned Role-Context Graph Policy for adaptive LLM-based multi-agent collaboration. The method jointly predicts agent participation, functional roles, and directed communication edges, using an inactive node state to adapt team size. The paper evaluates RCGP on six benchmarks covering general reasoning, mathematical reasoning, and code generation. It reports improvements over single-agent, static multi-agent, and automated topology-design baselines, as well as ablations, token-efficiency comparisons, robustness tests, and experiments with several LLM backbones.

Critical strengths
  1. The paper presents a reasonably coherent formulation that couples agent participation, role assignment, and communication topology in one graph-policy space.

  2. Modeling inactive agents explicitly is a sensible design choice, since team composition and communication structure are naturally interdependent.

  3. The reported ablation on the inactive category provides some evidence that jointly adapting participation and topology can be beneficial.

Critical weaknesses
  1. The main concern is the limited and potentially fragile evaluation protocol. RCGP is optimized separately for each benchmark, apparently using only 20 training queries for MMLU and 50 queries for each other benchmark. This setup raises substantial questions about selection bias, sensitivity to the training subset, and whether the learned policy generalizes beyond benchmark-specific optimization.

  2. The main comparisons use GPT-4o-mini, while the cross-backbone experiment compares RCGP only to Vanilla prompting rather than to competitive adaptive multi-agent baselines. Therefore, it does not establish that RCGP remains superior to prior topology-design methods across different backbones.

  3. The technical contribution appears to combine several plausible design choices—joint node/edge prediction, inactive nodes, directional role-context aggregation, structure-aware initialization, and RL—but the paper does not convincingly isolate what is fundamentally new or necessary beyond existing graph-based multi-agent design frameworks.

  4. The Automated Multi-Agent System Design subsection is overly enumerative. It introduces a large number of closely related methods, but only a small subset is included in the empirical comparison. The paper should better explain the relationship between the proposed method and these prior approaches, and clarify why certain highly relevant methods are excluded from the experiments.

  5. The manuscript uses dense terminology such as “role-context,” “shared-source,” and “shared-target” relations; clearer intuitive examples or a running example would improve accessibility.

  6. Figure 3 communicates the intended performance–cost trade-off, but it is visually dense and would be more convincing with error bars, consistent scaling, and a clearer definition of whether reported token counts include optimization/training costs or only inference costs.

  7. The paper should also ensure that all figures remain readable in standard two-column print size.

Questions for authors

See above, and one more question: Are all compared automated multi-agent methods given equivalent LLM-call budgets, token budgets, optimization steps, candidate team-size ranges, and access to the same role/template library? If not, please provide a budget-matched comparison.

If the authors can address my concerns well, I will be willing to increase the score.

7 comments

We appreciate the reviewer’s constructive comments and have revised the manuscript to include these clarifications.

Response to W1:

We follow the evaluation protocol used by our baselines G-Designer (Zhang et al., 2025d) and RADAR (Zhang et al., 2026b). We randomly select 20 training queries for MMLU and 50 for each other benchmark, optimize the policy on these subsets, and freeze it for evaluation on held-out queries. These training sizes are comparable to the baselines’ settings of 12--40 queries on MMLU and 40--50 on the other benchmarks. We keep the randomly selected subsets fixed across runs, vary the random seeds, and report the mean and standard deviation over two independent runs to quantify training and inference variability, following prior work such as GPTSwarm (Zhuge et al., ICML 2024), ADAS (Hu et al., ICLR 2025), and AFlow (Zhang et al., ICLR 2025).

To further assess sensitivity to the choice of training queries, we re-sample the training subset with two random seeds on the two largest benchmarks, GSM8K and SVAMP, while keeping all other settings unchanged. As shown in the table below, RCGP shows little sensitivity to training-subset selection, maintaining stable accuracy and inference token consumption across different subsets.

MetricGSM8KSVAMP
Accuracy (%)94.05 ± 0.5094.21 ± 0.30
Tokens (×1e6)8.61 ± 0.054.67 ± 0.06

Additionally, our cross-benchmark transfer experiments assess generalization beyond benchmark-specific optimization. We train our RCGP and existing RADAR on GSM8K and transfer their frozen policies to SVAMP and AQuA without further optimization. As shown in the table blow, GSM8K-trained RCGP achieves 94.63% and 81.86%, respectively, outperforming GSM8K-trained RADAR, our strongest baseline, on both benchmarks. These results support cross-benchmark transfer within mathematical reasoning.

MethodGSM8K → SVAMPGSM8K → AQuA
RADAR93.1681.37
RCGP94.6381.86

Response to W2:

We conduct an additional backbone evaluation, comparing RCGP with existing RADAR and ARG-Designer, the two strongest baselines in our main experiments, on all six benchmarks. Since deepseek-V4-flash (the strongest backbone we used in the original paper) is no longer available, we use deepseek-V4.1-flash for all methods. As shown in the table below, RCGP outperforms both methods across all six benchmarks, with the largest relative improvements of 3.74% over RADAR on HumanEval and 2.47% over ARG-Designer on MMLU. These results support RCGP’s advantage over our strongest baseline beyond the GPT-4o-mini backbone.

MethodMMLUGSM8KSVAMPAQuAHumanEvalMBPP
ARG-Designer90.9897.1696.4289.7195.6193.37
RADAR92.4897.0196.5389.9593.8694.96
RCGP93.2397.3296.6390.2097.3795.49

Response to W3:

Our novelty lies in reformulating collaboration design as a unified policy decision over participation, roles, and communication.

First, we jointly adapt team composition and topology through parallel node–edge prediction by introducing an inactive node state that makes participation directly constrain executable communication. In contrast, existing fixed-team optimizers (GPTSwarm, G-Designer) leave participation outside topology optimization, while participation-adaptive methods select from predefined compositions (MaAS) or construct graphs sequentially (ARG-Designer).

Second, we further identify the functional asymmetry of role contexts and propose directional structural encodings to distinguish upstream information support from downstream contribution use, which existing direction-aggregated summaries such as RADAR’s obscure. We use execution feedback to refine these encodings and graph decisions jointly.

Overall, we enable joint, parallel adaptation of participation and communication guided by directional role context, a capability existing methods do not support.

Response to W4:

We will follow your suggestion and streamline the related-work subsection by organizing prior approaches around their optimization targets and graph-generation mechanisms, clarifying their relationship to RCGP. We will move broader background material to the appendix. Our baseline selection prioritizes the latest directly relevant work, particularly methods published in 2026, including RADAR (ICML 2026), ARG-Designer (AAAI 2026), and GTD (ACL 2026), covering both autoregressive and diffusion-based topology generation. We discuss CARD (Wu et al., ICLR 2026) as complementary work: it emphasizes topology adaptation to changes in models, tools, and knowledge resources, whereas our evaluation focuses on query-conditioned collaboration with model and resource configurations held fixed within each comparison. We will clarify this evaluation scope and baseline-selection rationale, and remain open to further comparisons under the same setting.

Response to W5:

We include an example using the initial context graph in Fig. 2 to clarify these terms. A shared-source relation connects agents that receive information from the same agent. In Fig. 2, the Economist and the downstream Math Analyst share the upstream Math Analyst as their source. A shared-target relation connects agents that send information to the same agent. In Fig. 2, the Economist, the upstream Math Analyst, and the Physicist share the downstream Math Analyst as their target. Role-context describes how roles are organized around an agent from receiving and sending perspectives. For the Economist, its receiving context includes the upstream Math Analyst and the Physicist as incoming neighbors and the downstream Math Analyst as a shared-source neighbor; its sending context includes the downstream Math Analyst as an outgoing neighbor and the upstream Math Analyst and the Physicist as shared-target neighbors. We summarize these neighborhoods through role diversity and relative neighborhood size.

Response to W6:

We revise Fig. 3 to improve readability and clarify the performance--cost comparison. As detailed in Section 5.3, Fig. 3 counts inference tokens only: prompt and completion tokens from all agent calls and the final aggregation call across the test set. It excludes training costs. Table 10 in Appendix E.5 reports total training and inference tokens, with training and inference times listed separately.

Response to W7:

We enlarge all figure fonts to at least the caption font size, split Fig.~2 into two panels, and check that all figures are readable at standard print size.

Response to Q1:

Our reported experiments already provide budget-matched comparisons: we evaluate all methods under the same LLM-call and token-budget upper bounds. We also use the same per-call output-token limit of 1,000 tokens and the same role/template library. Actual calls and token consumption vary with the generated collaboration graph, and we include the final aggregation call in the reported execution cost. Following RADAR (Zhang et al., ICML 2026), we use five agents for fixed-team methods and allow adaptive methods to select up to five agents per query. Optimization steps follow each method’s recommended settings. We explicitly state these shared budget limits and method-specific settings in Appendix C.

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.

Reviewer_j5aUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RCGP, a query-conditioned graph policy for adaptive LLM-based multi-agent collaboration. The method jointly optimizes agent participation, role assignment, and directed communication through a unified graph action space. It further introduces directional role-context representations to capture the different functions of incoming and outgoing communication, and employs execution-guided reinforcement learning to optimize collaboration structures. Experiments on six benchmarks show consistent improvements over existing multi-agent approaches, with additional gains in token efficiency.

Critical strengths

A coherent framework for adaptive collaboration. The proposed method integrates participation, role assignment, and communication decisions within a unified graph policy. The use of separate receiving and sending role-context representations is also an interesting design choice that provides additional structural information for graph prediction.

Consistent empirical improvements. The method achieves the best performance across all six evaluated benchmarks, outperforming several competitive baselines, including RADAR and ARG-Designer. The results are reasonably convincing in terms of overall effectiveness, and the additional analysis of token consumption suggests a favorable performance-efficiency trade-off.

Useful ablation studies and analysis. The authors investigate several important components, including adaptive participation, directional role-context encoding, structure-aware initialization, and reinforcement learning. These experiments provide supporting evidence for the proposed design choices.

Critical weaknesses

The methodological novelty is somewhat incremental. Most components, including graph-based collaboration, adaptive agent selection, role-aware structural representations, and execution-guided policy optimization, have been explored in related work. The main contribution appears to be their integration with directional neighborhood statistics. While this combination is reasonable, it is not entirely clear whether it represents a substantial methodological advance over existing graph-based multi-agent optimization methods.

The role-context representation may be too coarse to capture meaningful collaboration semantics. The proposed encodings primarily rely on neighborhood role entropy and relative neighborhood size. These statistics characterize structural diversity but do not explicitly capture the relevance, quality, or actual information content of exchanged messages. Different communication structures could have similar role-context statistics despite having substantially different functional behaviors. This raises questions about how much additional information the proposed representation provides beyond standard graph attention and role embeddings.

The generalization of the learned graph policy remains insufficiently explored. The graph policy is optimized separately for each benchmark, using a small subset of queries from that benchmark. As a result, the experiments mainly demonstrate adaptation to unseen queries within the same benchmark rather than generalization across tasks or domains. It would be more convincing to evaluate whether a graph policy trained on one task can transfer to another without additional optimization.

Questions for authors

How does the proposed directional role-context encoding compare with more expressive alternatives, such as graph neural networks or attention-based representations that directly incorporate role identities and message content? Would such alternatives provide similar benefits without explicitly computing role entropy and neighborhood size?

Since the graph policy is trained separately for each benchmark, have the authors considered cross-benchmark transfer experiments? For example, how would a policy trained on GSM8K perform on SVAMP or AQuA without further optimization?

Could the authors report standard deviations or confidence intervals across independent training and inference runs? This would be particularly helpful for assessing the smaller improvements over RADAR on the code generation benchmarks.

4 comments

We thank the reviewer for the valuable feedback and have incorporated these clarifications into the new version of the manuscript.

Response to W1:

Our novelty lies in reformulating collaboration design as a unified policy decision over participation, roles, and communication.

First, we jointly adapt team composition and topology through parallel node–edge prediction by introducing an inactive node state that makes participation directly constrain executable communication. In contrast, existing fixed-team optimizers (GPTSwarm, G-Designer) leave participation outside topology optimization, while participation-adaptive methods select from predefined compositions (MaAS) or construct graphs sequentially (ARG-Designer).

Second, we further identify the functional asymmetry of role contexts and propose directional structural encodings to distinguish upstream information support from downstream contribution use, which existing direction-aggregated summaries such as RADAR’s obscure. We use execution feedback to refine these encodings and graph decisions jointly.

Overall, we enable joint, parallel adaptation of participation and communication guided by directional role context, a capability existing methods do not support.

Response to W2/Q1:

Regarding message relevance, quality, and content, these are properties observed during execution, whereas the task here is to generate the topology that determines that execution. Accordingly, common formulations, including existing G-Designer, RADAR, ARG-Designer, and our approach, use query and agent/graph information as generator inputs during both training and inference, without requiring actual exchanged messages. We explicitly encode role identities through node embeddings, together with edge representations (Eq. 8). Our directional role-context encoding supplements these representations with structural information rather than replacing them with entropy and neighborhood size alone. Task-level execution feedback trains the policy to use this information for effective collaboration.

To assess the effectiveness of our directional role-context encoding, we conduct two variants that replace it with a GAT-based or GCN-based encoder, following the reviewer’s suggestion. As shown in the table below, our encoding achieves relative improvements in average score of 3.09% and 3.47%, respectively, across the three benchmarks. These results support explicitly encoding role diversity and neighborhood support from separate receiving and sending perspectives, rather than relying entirely on implicit aggregation. The directional ablation further supports this distinction: merging receiving and sending contexts causes a 4.77% relative performance drop on HumanEval. Our encoding thus provides a useful structural inductive bias that complements role identity and edge representations using information available before execution.

VariantMMLUHumanEvalAQuA
RCGP w/ GAT81.9589.4782.35
RCGP w/ GCN81.2086.8484.80
RCGP84.2192.1185.29

Response to W3/Q2:

Following the reviewer’s suggestion, we evaluate cross-benchmark generalization by training our RCGP and existing RADAR on GSM8K and transferring their frozen policies to SVAMP and AQuA without further optimization. As shown in the table below, our GSM8K-trained RCGP policy achieves 94.63% on SVAMP and 81.86% on AQuA, outperforming the GSM8K-trained RADAR policy, our strongest baseline, on both benchmarks. These results support the transferability of the learned collaboration policy across mathematical reasoning benchmarks. On SVAMP, it also surpasses target-trained RCGP (94.00%) and RADAR (93.26%). The frozen policy remains query-adaptive: it determines agent participation, roles, and communication for each target query, reusing learned collaboration patterns without fixing the generated graph.

MethodGSM8K → SVAMPGSM8K → AQuA
RADAR93.1681.37
RCGP94.6381.86

Response to Q3:

As stated in Appendix C, we already average results over two runs with different random seeds and fixed training and test subsets, following prior work such as GPTSwarm (Zhuge et al., ICML 2024), ADAS (Hu et al., ICLR 2025), and AFlow (Zhang et al., ICLR 2025). Due to limited table space, we report the mean and the improvement over Vanilla in the main table of the paper. Following the reviewer’s suggestion, we now additionally report the standard deviation across two independent runs alongside the mean.

MethodsMMLUGSM8KSVAMPAQuAHumanEvalMBPPAvg.
Vanilla78.20 ± 0.0087.47 ± 0.1186.74 ± 0.1578.92 ± 0.6987.72 ± 0.0084.62 ± 0.0083.94 ± 0.11
CoT78.95 ± 1.0687.16 ± 0.3387.37 ± 0.7475.49 ± 0.0088.60 ± 1.2486.21 ± 0.7583.96 ± 0.03
ComplexCoT79.70 ± 1.0686.92 ± 0.5687.68 ± 0.6075.98 ± 0.6987.72 ± 0.0081.17 ± 1.1383.19 ± 0.49
MultiPersona79.32 ± 1.5987.55 ± 0.4587.05 ± 0.4579.41 ± 0.6988.16 ± 1.8676.39 ± 0.7582.98 ± 0.11
LLM-Debate78.57 ± 0.5389.52 ± 0.3389.05 ± 0.6079.90 ± 0.6989.04 ± 0.6282.23 ± 0.3884.72 ± 0.21
AgentVerse80.45 ± 0.0089.91 ± 0.2288.42 ± 0.3077.94 ± 0.6989.47 ± 0.0079.58 ± 1.1384.30 ± 0.22
MacNet81.95 ± 0.0088.02 ± 0.6786.00 ± 0.4579.17 ± 1.0485.09 ± 0.0084.88 ± 0.7584.19 ± 0.11
GPTSwarm76.69 ± 0.0089.20 ± 0.2288.74 ± 0.6080.88 ± 0.0089.47 ± 0.0079.58 ± 0.7584.09 ± 0.06
G-Designer80.83 ± 0.5391.10 ± 0.1190.00 ± 0.0080.64 ± 0.3589.91 ± 0.6288.33 ± 1.1386.80 ± 0.42
AgentPrune80.08 ± 0.5391.96 ± 0.1190.42 ± 0.4581.37 ± 0.6987.72 ± 0.0086.74 ± 0.0086.38 ± 0.08
GTD80.45 ± 0.0091.41 ± 0.2290.32 ± 0.7480.15 ± 1.0488.60 ± 1.2484.08 ± 0.7585.83 ± 0.42
MaAS81.20 ± 0.0091.17 ± 0.5689.68 ± 0.7480.39 ± 0.0090.35 ± 0.0085.41 ± 0.3886.37 ± 0.28
ARG-Designer81.58 ± 1.5991.25 ± 0.6792.21 ± 0.4581.62 ± 0.3589.47 ± 1.2487.00 ± 0.7587.19 ± 0.37
RADAR81.95 ± 1.0692.51 ± 0.0093.26 ± 0.7482.84 ± 0.6990.79 ± 0.6289.12 ± 0.3888.41 ± 0.23
Ours84.21 ± 0.0094.01 ± 0.1194.00 ± 0.1585.29 ± 0.6992.11 ± 0.0089.92 ± 0.3889.92 ± 0.05

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.

Dear Reviewer,

We sincerely apologize for the additional reminder. With only about six hours remaining in the discussion period, we would like to kindly follow up on our previous message.

We would greatly appreciate it if you could take a moment to review our responses. If you have any remaining concerns, we would be happy to address them. If our clarifications have resolved your concerns, we would be grateful if you could reconsider your rating.

Thank you again for your valuable time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.