acceptodds
Under review as a conference paper at ICLR 2027

FedGOTA: Towards Effective Federated Multi-Agent Collaboration via Graph Optimal-Transport Alignment

Abstract

In multi-agent systems (MAS) powered by large language models (LLMs), communication topologies orchestrate interactions among specialized agents, enabling coordinated reasoning on complex tasks. However, learning these topologies is often coupled with task supervision, making the resulting structures task-specific and difficult to transfer to new tasks. In federated settings, this coupling can further confine clients to task-specific knowledge silos, hindering effective aggregation and cross-client knowledge transfer. To address these limitations, we decouple structure learning from task supervision by introducing a shared structural self-supervised objective. Building on this formulation, we propose FedGOTA, a federated multi-agent collaboration framework via graph optimal-transport alignment. On each client, several complementary collaboration views are fused to train a topology generator with the self-supervised objective, allowing it to produce reusable, question-aware collaboration structures without task labels. At the server, we probe client topology generators with shared semantic anchors to construct graph codebooks, whose optimal-transport alignment guides personalized aggregation to mitigate negative transfer across heterogeneous clients. Extensive experiments on six reasoning benchmarks across four task domains show that FedGOTA achieves competitive performance compared with fourteen baselines spanning single-agent reasoning and multi-agent collaboration.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Federated Learning, Multi-Agent Collaboration, Graph Structure Learning.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tnbpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the "knowledge silo" and negative transfer issues in LLM-based Multi-Agent Systems (MAS) under federated learning settings, where topology learning is tightly coupled with task-specific supervision. To this end, the authors propose FedGOTA. On the client side, it employs multi-view graph self-supervised contrastive learning to decouple topology generation from task labels. On the server side, it probes client generators with shared semantic anchors to construct graph codebooks and introduces Optimal Transport (OT) alignment to quantify structural affinity among clients, thereby enabling personalized parameter aggregation. Experiments on six cross-domain reasoning benchmarks demonstrate that FedGOTA outperforms existing baselines in both reasoning accuracy and token efficiency. Ablation studies and other analytical experiments further validate the rationale behind the FedGOTA design.

Critical strengths
  1. Clear Presentation and Motivation. The paper is well-written, and the motivation for addressing the coupled topology learning and task supervision is logically sound and easy to follow. The transition from the identified bottlenecks to the proposed methodology is smooth.
  2. Solid Theoretical Formulation. The theoretical foundation is rigorous. The formalization of self-supervised decoupling and the structural affinity measurement based on Optimal Transport are well-designed and theoretically grounded.
  3. Comprehensive Empirical Evaluation. The experiments cover six authoritative benchmarks across four domains and compare against 14 baselines, including recent adaptive MAS techniques. Ablation studies, token consumption analysis, aggregation weight visualization, and case studies mutually corroborate the effectiveness of the proposed method.
Critical weaknesses
  1. Lack of Direct Empirical Validation for Core Claims. The central premise of the paper is that self-supervised decoupling enables the reuse and transfer of cross-task collaboration structures, and OT alignment facilitates personalized cross-client knowledge transfer. However, the current experiments do not directly validate these claims. For instance, it remains unclear how a generator trained on Task A performs when directly applied to generate topologies for Task B. Although the paper repeatedly emphasizes "reusability," it fails to introduce any unseen tasks during training to verify the generalization ability of the generator.
  2. Insufficient Justification for the Optimal Transport (OT) Module. As a core innovation highlighted in the title, OT is not compared against simpler structural distance metrics, such as Cosine Similarity between adjacency matrices, Frobenius norm, or Maximum Mean Discrepancy (MMD). If these computationally cheaper metrics can yield similar aggregation weight distributions and final performance, the necessity of introducing the heavy OT machinery becomes questionable.
  3. Limited Efficiency and Scalability Analysis. The efficiency analysis is currently restricted to theoretical complexity in the appendix, lacking empirical data on actual wall-clock time and memory footprint. There is no comparison of the actual training/inference time between FedGOTA and baseline methods. Furthermore, to justify the use of OT, it is crucial to compare the time and memory consumption of the OT module against simpler distance metrics. Without this, the practical advantage of OT is not fully supported.
Questions for authors
  1. Generalization Experiments: Regarding Weakness 1, please provide zero-shot cross-task transfer experiments (e.g., applying a generator trained on Task A directly to Task B) or evaluations on completely unseen tasks. This is critical to substantiating the paper's core claims regarding "reusability."
  2. OT vs. Simpler Metrics: Regarding Weakness 2, please provide ablation studies comparing the OT alignment with simpler structural distance metrics (e.g., Frobenius norm, Cosine similarity, MMD) in terms of both final performance and the resulting aggregation weight distributions.
  3. Empirical Efficiency Analysis: Regarding Weakness 3, please provide quantitative analyses of wall-clock time consumption and peak memory footprint for both FedGOTA and representative baselines. Specifically, how much overhead does the OT module add compared to simpler distance metrics?
  4. Performance Gap on MMLU-Pro: Could the authors elaborate on why FedGOTA underperforms compared to specific baselines on the MMLU series, particularly the >1% gap on MMLU-Pro? Is this related to the shared structural prior or the nature of general reasoning tasks?
3 comments

For W1, in our federated setting, clients keep their data local and learn topology generators without task-specific answer supervision. During inference, each client directly employs its personalized generator without further fine-tuning, enabling zero-shot inference under its local task distribution. Moreover, the local topology generator introduced in Section 4.1 (hereafter referred to as FedGOTA-L) enables direct cross-task reuse without task-specific supervision or further optimization. As shown in Table 2 and Table 3, FedGOTA-L trained on MMLU and MMLU-Pro or GSM-H and AQuA can be directly transferred to unseen tasks without further fine-tuning, consistently outperforming Vanilla and OFA-MAS. Furthermore, as shown in Table 4, a single FedGOTA-L foundation model jointly trained across all pooled datasets without the federated constraint achieves a 71.13% average accuracy when reused across all tasks, surpassing all baselines on average.

|Table 2 (Source: MMLU + MMLU-Pro) | AQuA | GSM-H | HumanEval | | -- | - -| -- | -- | | Vanilla | 81.28% | 47.95% | 82.27% | | OFA-MAS | 79.00% | 48.85% | 87.94% | | FedGOTA-L | 84.02% | 61.64% | 88.65% |

Table 3 (Source: GSM-H + AQuA)MMLUMMLU-ProGPQA
Vanilla74.43%54.34%39.73%
OFA-MAS82.65%58.45%34.70%
FedGOTA-L84.47%62.10%42.47%
Table 4MMLUGSM-HHumanEvalAQuAMMLU-ProGPQAAVG
AgentDropout84.32±1.6058.45±2.3786.76±1.0883.11±0.9259.06±2.3438.81±1.9968.42%
G-Designer81.59±0.5256.01±1.2685.34±1.6482.95±1.6056.32±1.8740.94±1.2967.19%
ARG-Designer83.17±1.0449.62±1.9085.11±1.4281.28±1.6563.83±2.2539.12±1.4067.02%
OFA-MAS81.73±1.2148.10±0.7086.53±0.7181.13±2.3058.14±0.9537.29±1.8565.49%
FedGOTA-L84.63±0.7061.19±0.4689.15±0.4384.93±0.4663.17±0.9543.71±1.3671.13%

For W2, we have conducted additional ablation experiments comparing OT alignment with simpler affinity metrics, including Graph Cosine, Weight Cosine, Frobenius norm, and MMD. As shown in Table 5, OT consistently achieves competitive performance across all benchmarks.

Table 5MMLUMMLU-ProGPQAGSM-HAQuAHumanEvalAVG
Graph Cosine83.26±0.5360.43±1.4740.64±2.0960.58±1.1581.43±1.8587.94±1.2369.05
Weight Cosine83.87±1.4060.43±1.4040.49±1.4057.69±1.4780.97±0.7088.89±2.0568.72
Frobenius83.41±1.4060.73±1.3739.27±1.5859.97±2.6882.34±0.7088.65±1.8869.06
MMD83.11±0.4660.58±0.9541.10±1.2160.58±1.0581.74±1.2188.65±0.7169.29
OT84.02±0.5061.34±0.7044.28±1.6062.41±1.9085.08±0.3088.91±1.5071.01

For W3, we quantitatively compare the alignment runtime and memory usage of different server-side affinity measures in Table 6. Although iterative OT takes longer than simpler affinity metrics, its server-side alignment requires only 3.44 seconds, with an average memory usage of 617.78 MiB.

Table 6Time (s)Average RSS (MiB)
OT3.439553617.78
Graph Cosine0.008370531.45
Weight Cosine0.025523556.82
Frobenius0.004812536.68
MMD0.021414542.79

For W4, the performance gaps on MMLU and MMLU-Pro in Table 1 should be considered in light of the different training objectives, as several adaptive baselines rely on task-specific supervision or LLM feedback to optimize collaboration structures. In contrast, FedGOTA learns collaboration topologies without task-specific supervision while keeping client data local. Through personalized federated learning across heterogeneous tasks, FedGOTA learns reusable collaboration topology generators that can be directly applied to downstream questions without further fine-tuning. Regarding the shared structural prior and general reasoning tasks, the performance gap on MMLU-Pro may partly reflect the trade-off between cross-task transferability and task-specific specialization. Specifically, FedGOTA learns transferable collaboration patterns across heterogeneous tasks, which may be less tailored to the complex reasoning demands of MMLU-Pro than methods optimized using feedback specific to each task. While supervised baselines may benefit from task-specific optimization on challenging reasoning tasks, FedGOTA emphasizes reusable collaboration patterns across heterogeneous tasks without relying on task-specific supervision.

Most of my concerns have been resolved. While the rebuttal slightly conflates "model generalization" with "federated marginal gains" (W1), and leaves the O(E3)O(E^3) scalability issue for larger agent pools unaddressed. Overall, I recommend upgrading the score to "Weak accept".

AuthorsOpenReview ↗

Thank you very much for your positive feedback and for recognizing the value of our work.

Thank you very much for your positive feedback and for recognizing the value of our work. In our rebuttal, Tables 2 and 3 focus on demonstrating cross-domain generalization, whereas Table 4 highlights that our method achieves results comparable to centralized joint training through federated collaboration, aiming to emphasize the effectiveness of federated learning. It is worth noting that multi-domain joint training indeed requires greater agent diversity and a larger agent pool compared to single-domain training methods such as G-Designer. However, compared to OFA-MAS, which relies on 21 candidate agents, our method remains highly competitive, utilizing a candidate pool of only 10 experts to achieve effective collaboration in a distributed setting without sharing raw data.

Reviewer_8wYaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FedGOTA, a federated multi-agent collaboration framework through graph optimal transport alignment. Specifically, it trains a question-aware topology generator for LLM multi-agent systems on each client via a contrastive loss that matches each expert's GNN embedding under the learned graph with its embedding under a fixed DAG shared among all clients. Afterward, the server runs each client’s generator on shared “semantic anchor” questions to build graph codebooks. It compares those codebooks with entropic OT and uses the distances as softmax weights for personalized parameter averaging. The authors conduct experiments over six reasoning benchmarks across four task domains to show the competitive performance of FedGOTA compared with fourteen baseline methods.

Critical strengths
  1. Separating topology learning from task rewards is a well-motivated goal. Training the generator needs no LLM calls, which is a real cost advantage in multi-agent collaboration system design.
  2. The experiments are comprehensive, including the main performance comparison, cost analysis, ablation study, aggregation scaling, and prior topology analysis, etc.
Critical weaknesses
  1. The federal learning component contributes little, according to Figure 8. In particular, uniform aggregation achieves 67; full local training with no cross-client sharing (scale →∞\rightarrow \infty ) achieves 70.83, with the best setting at 71.00 at scale →2\rightarrow 2. To my understanding, the local-only variant still uses the same self-supervised topology training and expert selection, but receives nothing from other clients, making the OT affinity step irrelevant. Therefore, the performance gain in FedGOTA seems to come from local components, not from the two claimed contributions, i.e., the federated formulation and OT-guided personalized aggregation.

  2. The proposed method appears to be an incremental combination of existing techniques. On the server side, FedGOTA probes client models with shared inputs, measures how similar the clients are, and averages parameters with similarity-based weights. FED-PUB [1] and pFedGraph [2] already use this pattern for personalized federated learning, and optimal transport has already been applied in federated settings [3]. The main difference is that FedGOTA measures client similarity with OT over graph codebooks. The paper does not explain the rationale that the choice is preferable to simpler similarity measures.

[1] Personalized Subgraph Federated Learning. ICML'23.

[2] Personalized Federated Learning with Inferred Collaboration Graphs. ICML'23

[3] FedOTP: Global and Local Prompts Cooperation via Optimal Transport for Federated Learning. CVPR'24.

Questions for authors

See above.

2 comments

For W1, we would like to clarify that the default setting of Scale\text{Scale} is 1 (Fig. 4), under which FedGOTA achieves an average accuracy of 72.33%, rather than the 71.00% reported for Scale=2\text{Scale}=2 in Appendix Fig. 8 as an additional comparison. Compared with independent local training, personalized federated aggregation yields an improvement of 1.50 percentage points. In contrast, uniform aggregation suffers from negative transfer across heterogeneous tasks, resulting in a 3.83-percentage-point accuracy drop relative to independent local training. These results jointly support both core contributions of our work: independent local training demonstrates the effectiveness of client-side multi-view self-supervised graph learning without task-specific supervision, while server-side OT-guided personalized aggregation mitigates the negative transfer of uniform aggregation and enables positive cross-client knowledge sharing. These findings further highlight the complementary roles of client-side self-supervised learning and server-side OT-guided personalized aggregation. To further demonstrate the effectiveness of federated collaboration, Table 1 compares independent local training with FedGOTA across all six benchmarks, showing an improvement in average accuracy from 67.74% to 71.48%. Note that the above results are based on the first 100 evaluation questions from each benchmark, whereas Table 1 compares FedGOTA with independent local training on the full evaluation subsets.

Table 1LocalFedGOTA
MMLU81.28%84.47%
MMLU-Pro59.82%62.10%
GPQA36.07%43.84%
GSM-H58.45%63.01%
AQuA82.19%85.39%
HumanEval88.65%90.07%
Avg67.74%71.48%

For W2, FED-PUB, pFedGraph, and FedOTP primarily address personalized subgraph federated learning, client collaboration graph inference, and global and local prompt cooperation, respectively, under shared task objectives, rather than breaking down isolated collaboration knowledge silos across heterogeneous tasks in multi-agent systems (MAS). Moreover, in cross-task heterogeneous MAS settings, probing clients with shared inputs and aggregating their parameters using conventional similarity measures may fail to capture reliable structural affinities for knowledge transfer. As discussed in Section 4.3, clients with different tasks may exhibit similar reasoning and collaboration preferences through different but semantically related expert interaction edges. Conventional similarity measures do not explicitly account for semantic correspondences between different expert interactions, potentially overlooking structural affinities across heterogeneous expert roles. To address this limitation, we explicitly construct a ground transport cost matrix G(i,j),(u,v)G_{(i,j),(u,v)} (Eq. 21) based on the semantic differences between source and target experts, and solve an entropy-regularized optimal transport problem over graph codebooks (Eq. 22), enabling reliable structural affinity modeling across heterogeneous tasks. Therefore, multi-view self-supervised topology learning on clients and optimal transport alignment of graph codebooks based on expert semantics on the server are jointly designed to enable collaboration knowledge transfer across heterogeneous tasks, rather than a straightforward combination of existing methods.

Dear Reviewer 8wYa,

Thank you for your valuable insights and the time spent reviewing our work, especially during this busy period. We greatly appreciate it.

As we near the final day of the discussion period, we wanted to check in and see if our responses have addressed your concerns. We welcome your further queries and suggestions and are ready to provide any additional information or clarification you might need.

We understand that reviewing is a time-consuming task and greatly appreciate your commitment to advancing academic discourse. Our team is dedicated to improving our work based on your feedback, believing it will enhance our contribution to the field.

Thank you once again for your guidance and support.

Best regards, The Authors

Reviewer_mMK6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FedGOTA studies personalized federated multi-agent collaboration. It reformulates adaptive topology learning as a shared graph self-supervised objective, so that topology generators can be trained without task-specific supervision. For personalized aggregation, it uses server-side graph codebooks and optimal transport to estimate structural affinity across clients. On six reasoning benchmarks with GPT-4o-mini, the authors report an average accuracy of 71.01%, higher than fourteen baselines, with token consumption comparable to them.

Critical strengths

S1. Learning collaboration topologies without task-specific rewards makes the learned structure reusable across clients with different tasks, which is what makes the federated setting meaningful here.

S2. The evaluation is reasonably comprehensive, covering six benchmarks spanning general, mathematical, code, and scientific reasoning, and comparing against fourteen baselines.

S3. The aggregation-weight visualization and the prior-topology comparison are most useful, showing that clients with related tasks receive higher mutual weights.

Critical weaknesses

W1. The federated setting does not match the paper’s own motivation. Each of the six clients corresponds to one benchmark, with 20 local training instances, and training runs for two communication rounds. The introduction (C2 and Fig. 1) argues that clients may prefer different collaboration patterns even when solving the same task, yet no experiment places multiple clients on the same task or splits a task across clients. As evaluated, the setting is closer to multi-task learning with one task per client than to the federated scenario the paper motivates.

W2. The source of the gain is unclear. The training objective has no task signal; question awareness comes mainly from frozen BERT similarity in the task-affinity view (Eq. 10) and expert selection (Eq. 27). Fig. 11 suggests that with 10 experts, only about 20% or fewer questions receive a distinct topology. A control with the same expert selection but a fixed topology is needed to show that the learned topology matters.

W3. Federation adds little. In Fig. 8, local-only training reaches 70.83% and the best personalized aggregation 71.00%, far below the 2.6-point margin over AgentDropout. There is also no comparison of OT against simpler affinity measures, and the weights in Fig. 4 are mostly diagonal with weak cross-client differentiation.

W4. The self-supervised objective is not well justified. It aligns representations with a fixed random DAG, which carries no information about collaboration quality. Since the prior is the anchor of the loss, the “w/o S^p” ablation is underspecified, and its large drop may reflect losing the training signal rather than the value of structural guidance.

Questions for authors

Please refer to the above comments.

3 comments
AuthorsOpenReview ↗

We thank the reviewer for the constructive comments and look forward to further discussion.

For W1, assigning different benchmarks to individual clients represents a more challenging cross-task heterogeneous FL setting. Compared with conventional same-task FL, which primarily addresses data distribution heterogeneity, our setting additionally involves heterogeneous task requirements and collaboration preferences, complicating cross-client transfer and increasing negative transfer risks. C2 and Fig. 1 illustrate structural heterogeneity even within the same task domain. Specifically, our six client benchmarks span four task domains, with MMLU and MMLU-Pro representing related but distinct reasoning tasks. Such differences motivate structural affinity modeling across heterogeneous clients. We will clarify C2 and Fig. 1 accordingly.

For W2, our self-supervised objective aims to break task-specific knowledge silos by focusing on collaboration between queries and agents. This enables heterogeneous clients to learn under a unified objective. Question awareness extends beyond static BERT similarities by jointly modeling question semantics and structural priors. Specifically, learnable asymmetric projections capture question-conditioned directed expert interactions, which are then fused with semantic affinity and the shared structural prior. Through message passing in the graph encoder and cross-view contrastive learning, the topology generator learns question-aware collaboration structures. During inference, expert selection is jointly determined by the learned topological connectivity and semantic affinity.

In Fig. 11, the 20% unique ratio under 10 experts (B=5B=5) may be attributed to the fact that same-domain questions share similar reasoning patterns and naturally converge to stable core collaboration structures. Nevertheless, the substantial structural deviation from the shared prior (0.4–0.8) and the increase in unique topologies under 20 experts (up to 56%) demonstrate that FedGOTA learns structures distinct from the prior and exhibits greater question-dependent structural diversity as the expert space expands.

Regarding the fixed-topology comparison, Appendix A.5 (Fig. 9) shows that FedGOTA generally outperforms directly executing various fixed prior topologies under the same candidate expert pool, although the selected experts may differ between the two settings. Moreover, Fig. 7 shows that edge pruning with fixed expert selection generally reduces accuracy, while replacing the GNN with an MLP lowers average accuracy from 71.48% to 67.25%, supporting the effectiveness of learned collaboration structures.

For W3, under the default setting (Scale=1, Fig. 4), FedGOTA achieves 72.33% average accuracy, compared with 71.00% at Scale=2 (Fig. 8), outperforming independent local training (70.83%) by 1.50 percentage points. While strong local performance (70.83%) supports the effectiveness of self-supervised graph learning, uniform aggregation suffers a 3.83-point drop, consistent with cross-task negative transfer. In contrast, OT-guided personalized aggregation achieves a 5.33 point gain over uniform aggregation. More importantly, personalized federated learning enables cross-task knowledge sharing while preserving client-specific collaboration preferences, which independent local training cannot achieve. Moreover, Fig. 4 (Scale=1) shows self-weights of 0.310–0.366, indicating that 63.4%–69.0% of aggregation weights come from other clients, with stronger affinities between related tasks (e.g., MMLU and MMLU-Pro). Furthermore, we conducted additional ablation experiments comparing OT with simpler affinity metrics, including Graph Cosine, Weight Cosine, Frobenius, and MMD, as detailed in our response to Reviewer 8wYa (W2).

For W4, following MacNet, the fixed random DAG does not directly encode task-specific collaboration quality, but provides an acyclic structural bias and a shared, task-agnostic reference for self-supervised learning across heterogeneous clients. Instead, FedGOTA integrates the prior with learned question-aware interactions and semantic affinity, aligning the fused and prior graph representations through contrastive learning. Appendix A.5 (Fig. 9) further shows that FedGOTA generally outperforms direct execution of different structural priors (Chain, Star, Tree), suggesting that its effectiveness is not limited to the random DAG. Furthermore, the substantial structural deviation from the shared prior (0.4–0.9) indicates that FedGOTA learns distinct topologies rather than simply replicating the fixed prior. Regarding w/o Sp\mathbf{S}^{p} in Fig. 5, removing the structural prior does not eliminate the contrastive signal, as the remaining views are contrastively aligned before fusion. The performance drop further suggests that the shared structural prior provides complementary guidance beyond the two question-dependent views.

Thanks for the detailed clarification. My concerns are well addressed, and I'd like to raise my score.

Thank you very much for your positive feedback and for raising your score. We truly appreciate the time and constructive insights you shared throughout the discussion, which have helped us significantly strengthen our work.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their constructive feedback and recognition of our work. In this paper, we explore an interesting topic: federated multi-agent collaboration under client heterogeneity, where FedGOTA decouples topology learning from task rewards without LLM calls during training or sharing raw data.

Despite our efforts to engage all reviewers, Reviewer 8wYa has not provided further feedback. In response to our rebuttal, two reviewers confirmed their concerns were resolved and raised their scores.

For the first concern W1 from Reviewer 8wYa, the reviewer mistakenly treated Scale = 2 at 71.00% in Appendix Fig. 8 as our default setting, whereas our actual default Scale = 1 in Fig. 4 achieves 72.33% versus 70.83% for local training. Moreover, full-subset comparisons in Rebuttal Table 1 show FedGOTA improving average accuracy over local training from 67.74% to 71.48%.

For the second concern W2 from Reviewer 8wYa, classic FL algorithms FED-PUB, pFedGraph, and FedOTP address shared task objectives rather than cross-domain multi-agent collaboration. Beyond server-side personalization, FedGOTA shares no algorithmic similarity with FED-PUB or pFedGraph, and differs completely from FedOTP in optimization objectives and modeled expert nodes.

All additional experiments and clarifications have been included in our responses. We appreciate your consideration of this evidence in the final assessment.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.