acceptodds
Under review as a conference paper at ICLR 2027

Topology as Experience: Federated Optimization on Collaborations in LLM-based MAS

Abstract

Topology design for multi-agent systems built on large language models is typically optimized within individual clients, leaving useful collaboration experience from other clients underused. Sharing textual experience introduces growing records and the risk of exposing sensitive task content, while heterogeneous task distributions require client-specific collaboration strategies. To address these limitations, we propose a personalized federated framework that represents collaboration experience as a compact probabilistic topology graph. Specifically, each client combines a shared probabilistic graph that aggregates reusable communication patterns with a personalized probabilistic graph that adapts them to local tasks, while keeping raw tasks and execution traces local. To reduce the cost of topology optimization, a task-conditioned mixture-of-experts predictor learns from historical execution records and supplies inexpensive rewards for probabilistic graph training. Limited real multi-agent feedback then refines the topology policy through reinforcement learning to address the gap between predicted and executed performance. At deployment, the predictor ranks sampled topologies and selects one for execution. Across two heterogeneous three-client suites spanning coding, mathematics, and reasoning, the complete framework achieves the highest mean accuracy in both suites. Predictor pretraining improves the starting topology policy, limited real feedback further refines it, and personalized candidate search reaches 90.0% accuracy with eight candidates in a setting where the shared policy requires 512. These results show how a probabilistic graph turns cross-client collaboration experience into executable, client-adaptive policies.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multi-Agent Systems, Large Language Models, Topology Optimization, Federated Learning, Personalized Learning

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_Xq9EOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that existing multi-agent collaboration topology optimization methods are primarily developed independently within individual clients, making it difficult for clients to reuse collaboration experience from one another. To address this issue, the paper proposes a personalized federated topology optimization framework that represents collaboration experience as probabilistic parameters of communication edges, thereby constructing a general shared probabilistic graph. When transferring this shared knowledge to individual clients, the framework combines the shared graph with local parameters to generate personalized topologies. Additionally, the paper trains a performance predictor to evaluate candidate topologies.

Critical strengths
  1. The experiments are relatively comprehensive, covering different task types and ablation studies.

  2. The paper addresses a meaningful problem: the difficulty of reusing collaboration experience across different clients, which has practical value.

  3. The paper considers practical costs and designs a performance predictor to reduce the cost of topology optimization.

Critical weaknesses
  1. The transferability of the experience summarized in the shared probabilistic graph across different clients is not sufficiently justified. For example, in tasks requiring a definite decision, the topology involving decision-making roles may be important, whereas in open-ended tasks, such roles may be less necessary or even dispensable.

  2. The final performance improvements appear to be relatively small, and it is unclear whether these gains could be partly due to randomness.

  3. The precise definition of a “client” is not sufficiently clear, particularly regarding whether the proposed framework still applies when different clients use different large language models.

  4. The unique contribution of this work compared with existing methods is not sufficiently highlighted, making it difficult to identify the most distinctive insight of the paper.

Questions for authors

Thanks for your paper but I have some little problem

  1. Why should the experience summarized in the shared probabilistic graph be transferable across different clients? For example, in tasks requiring a definite decision, the topology involving decision-making roles may be very important, whereas in open-ended tasks, such roles may be less necessary or even dispensable.
  2. The final performance improvements appear to be relatively small. Could these gains be partly due to randomness?
  3. What is the precise definition of a “client” in this paper? Does the proposed definition and framework also apply when different clients use different large language models?
  4. Compared with existing methods, what is the unique contribution of this work? Which part of the paper best highlights its most distinctive insight?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Xq9E

We thank Reviewer Xq9E for raising the questions about transferability, randomness, the client definition, and the distinctive contribution.

Q1. Transferability. The shared object is routing structure rather than task content. The two suites use three non-IID clients covering coding, mathematics, and reasoning under a common role schema. Shared logits encode reusable entry and decomposition patterns; the personalized component adjusts edges for each client's task distribution, and the task-conditioned predictor ranks candidates per task. In Table 2, the complete configuration has the highest mean in both suites and improves over Shared + Predictor on five of six clients. The client-level variation is handled by sharing reusable topology structure while retaining client-specific edges.

Q2. Randomness. We evaluate this concern using matched outcomes rather than independent averages. Complete and Local + Predictor use the same pretrained predictor initialization, local-update rule, feedback schedule, held-out items, and fixed seed. On the same 309 items per suite, Complete wins in 19 vs. 6, 24 vs. 9, 16 vs. 5, and 10 vs. 2 item pairs for AF-3, GD-3, AF-20, and GD-20, respectively; the exact two-sided McNemar p-values are 0.014, 0.014, 0.027, and 0.039. The corresponding 95% paired intervals are all above zero: [+0.97, +7.45], [+1.29, +8.41], [+0.65, +6.47], and [+0.65, +4.85] points. The same direction appears in both workflow suites and at both local-data sizes. The 0.65/0.97-point margins in Table 2 compare with the strong CoT reference; the matched federated comparison is 2.59–4.86 points. Thus, the reported federated margin is a consistent paired contrast across the two suites and both local-data settings, rather than an isolated mean from one client or one test set.

Q3. Client. A client is a unit with its own task distribution, local history, and executor configuration. The main experiments use three non-IID clients per suite with GPT-4o-mini as the suite executor. Appendix G additionally transfers the predictor from GPT-4o-mini to Qwen3-8B using 240 target-executor outcomes while retaining the task splits, agent pool, topology policy, and role schema. Eqs. (5)–(6) aggregate aligned topology parameters rather than model weights, so executor heterogeneity is compatible when clients share the role schema and edge coordinates.

Q4. Distinctive contribution. The core contribution is to use aligned communication-edge logits as the federated experience unit and decompose them into a shared tensor and a client-specific tensor in Eqs. (5)–(6). This lets clients reuse executable routing structure through mean aggregation while preserving local preferences. Predictor guidance is the efficiency component that reduces the cost of searching these distributions. Figure 2 and Section 4.2 highlight this shared/personalized topology decomposition; compared with DAWN, the exchanged object is aligned edge logits rather than a workflow generator.

Thank you for your detailed response and additional experiments. I appreciate your efforts, particularly the additional statistical analysis addressing the concern about randomness.

However, I feel that my two main concerns have not been fully resolved.

  1. Transferability. I remain unconvinced that the collaboration experience summarized in the shared probabilistic graph is generally transferable. First, different task types may require substantially different collaboration structures. Second, different LLMs may exhibit different collaboration behaviors. Although your additional experiments, including the Qwen3-8B predictor-transfer experiment, provide some supporting evidence, they do not fully address these concerns or establish the conditions under which such experience can be effectively transferred.

  2. Distinctive contribution. Since the three main technical components have foundations in existing research, I see the paper's primary contribution as a system-level integration and exploration of these methods. I recognize the value of this combination and its practical design contribution. However, I am still not convinced that the paper has sufficiently demonstrated the distinctive advantages of sharing aligned edge parameters over closely related federated methods such as DAWN.

Given these remaining concerns, I would prefer to maintain my score and consider the perspectives of the other reviewers before deciding whether to raise my rating.

Reviewer_6N8XOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses probabilistic graphs to represent collaboration patterns in MAS and share them across clients through federated learning. The authors combine shared and personalized topology parameters to handle different client tasks, and use a pretrained predictor to reduce agent executions for topology optimization.

Critical strengths

The paper investigates a new way to share collaboration experience across distributed LLM-based MAS. The probabilistic graph representation is simple and practical, and the experiments provide some useful observations about personalization and predictor-guided topology selection.

Critical weaknesses
  1. The key motivation of this work is that clients can benefit from useful collaboration patterns learned by other clients. However, the experiments do not clearly demonstrate that exchanging topology parameters produces better policies than learning them independently under identical conditions. the baselines compared in Table 2 are mainly single-client MAS methods with different optimization strategies. So the improvements could come from the predictor, topology parameterization, or different search procedures rather than federation itself

  2. In Sec. 2 author mentioned DAWN, which is a parallel work that explicitly identifies the problem of aggregating conflicting communication patterns across clients. Key difference is that DAWN learns distributed workflow generators and uses SVD-based alignment and structural regularization, whereas this paper directly aggregates aligned edge logits. The authors need to explain what is gained or lost by replacing a learned workflow generator with directly aggregated probabilistic edges.

  3. From the main result table 2 in sec. 5.2, the actual improvement margins are very small. 0.65% for AF and 0.97% for GD

  4. This paper uses a task-conditioned predictor to guide topology optimization and to select a topology at deployment. But the predictor is centrally pretrained on historical workflow records. Its ability to transfer knowledge across task families may already account for much of the cross-domain benefit instead of federated topology updates

  5. The experimental setting is limited, as the authors evaluate only three clients per suite, and all clients share the same agent pool, graph schema, and aligned edge-parameter coordinates

  6. The authors claim that exchanging topology parameters is safer than exchanging textual execution records, but aggregated preferences could still reveal information about the types of tasks, tools, or workflows favored by individual clients. Although the author specify this in the conclusion as future work, considering this application field as a personalized FL framework, clarifying the potential risk and direction is very important

Questions for authors

See the weaknesses listed above

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6N8X

We thank Reviewer 6N8X for the questions about the federated/local comparison, DAWN, the performance margin, predictor control, the client setting, and privacy.

W1. Federated aggregation versus local learning. Table R1 adds matched Local only and Local + Predictor configurations without federated aggregation. The results separate two effects. Shared only − Local only is +2.91 AF/+1.94 GD when local data is insufficient (3 tasks) and −0.97/+0.32 when local data is sufficient (20 tasks), so the observed sharing margin is larger in the insufficient-data setting. Complete − Local + Predictor (without federated aggregation) is +4.21/+4.86 when local data is insufficient and +3.56/+2.59 when local data is sufficient, so the complete method remains ahead of its local counterpart in both settings. The first comparison isolates aggregation without predictor guidance; the second follows the paper's predictor protocol, with the same pretrained checkpoint and local update rule on both sides, and evaluates the complete configuration against its local counterpart.

For convenience, Table R1 is reproduced here as well:

ConfigurationAF 3AF 20GD 3GD 20
Local only74.7679.6155.0267.64
Local + Predictor (without federated aggregation)77.0280.9156.3170.55
Shared only (β=0)77.6778.6456.9667.96
Shared + Predictor78.6482.2058.9071.84
Complete (ours)81.2384.4761.1773.14

The paired McNemar tests for these matched comparisons are reported once in the oxjR response above.

W2. DAWN comparison. The methods make different design tradeoffs. DAWN shares a learned workflow generator and uses alignment and structural regularization. We share aligned edge logits, so aggregation is a mean in a common coordinate system and requires no generator retraining or per-client re-alignment. This keeps the exchanged update fixed-size and simplifies deployment, while the shared edge-logit representation does not itself encode context-dependent workflow branching. The public DAWN paper does not specify the implementation and several key hyperparameters needed for a numerical reproduction, so we compare the structural design directly.

W3. Performance margin. The 0.65/0.97-point margins in Table 2 compare the complete method with the strongest non-sharing reference. Table R1 gives the matched comparison to the local predictor baseline: the complete configuration is ahead by 2.6–4.9 points across the two local-data settings and suites. Thus the margin remains with the same predictor initialization, local update protocol, and training schedule.

W4. Predictor control. The predictor is pretrained once, with the same initial checkpoint and architecture across configurations, and is never aggregated. The Complete versus Local + Predictor (without federated aggregation) comparison uses the same local predictor-update rule, β, proxy/feedback schedule, and evaluation tasks; predictor updates remain local, while topology updates are averaged only in the Complete configuration.

W5. Three-client setting. The evaluation uses three clients with a common agent pool and aligned role coordinates. The method targets this shared-schema setting; disjoint role coordinates require a separate alignment mechanism. Table R1 varies local data availability within this setting.

W6. Privacy. The exchanged edge tensor contains no task text or execution trace, whereas experience exchange would transmit raw records. Edge statistics can still reveal task-family preferences. Clipping and Gaussian noise could provide natural protections for this channel.

Reviewer_6N8XOpenReview ↗

Response to Authors

Dear Authors,

Thank you for the authors’ responses. However, my main concerns remain. The DAWN comparison, predictor compond, limited client setting, and privacy risk are acknowledged but not addressed empirically.

The key contribution also remains incremental, as the framework directly averaging edge logits is a simple extension of existing topology design in MAS (probabilistic edge graphs) and personalized FL techniques (interpolation). Thus, I will maintain my score.

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

The paper addresses topology design in multi-agent systems, which remains optimized independently for each client and may overlook useful collaboration experience discovered by other clients. Although using textual experience is intuitive, removing task-specific content and managing the accumulation of experience remain challenging. The authors sidestep these difficulties by representing collaboration experience as probabilistic topology graphs: each client maintains a shared graph for reusable communication patterns and a personalized graph for local experience. A pretrained predictor guides probabilistic graph sampling to provide a promising search initialization, which is then refined using actual feedback. However, the evaluations do not establish the additional value of cross-client sharing. Neither the accuracy comparisons nor the efficiency studies demonstrate that sharing improves upon independent local optimization. Since cross-client experience transfer is presented as a central contribution, its main claimed benefit remains unsubstantiated.

Critical strengths

The paper addresses the potential inefficiency of repeatedly rediscovering high-level collaboration experience in multi-agent systems.

Critical weaknesses
  1. The experiments do not establish the benefits of sharing and personalization over independent local learning. Cross-client experience transfer is a central contribution, but Table 2 lacks a matched local-only variant. The direct comparison in Table A7 shows, at best, small gains on AF, while both Shared and Personalized perform worse than Local on GD. Although personalization improves over the shared policy in Table 2, this could reflect mitigation of negative transfer from sharing rather than benefits from other clients’ experience. The reported gains may instead arise from predictor guidance and probabilistic topology optimization. A matched local-only control is needed to establish whether sharing contributes positively and whether personalization improves upon independent local optimization.

  2. The 64× candidate-budget result is reported only for the GD Reason-majority client. Figure 4(b) evaluates only this client, without specifying the evaluation set or explaining its relationship to the main experiments. This scope is too narrow to establish the candidate-budget benefit across heterogeneous clients.

Questions for authors

What is the main technical contribution beyond probabilistic topology optimization and predictor guidance? If it is cross-client experience sharing, which experimental comparison validates its benefit over independent local optimization?

2 comments
AuthorsOpenReview ↗

Response to Reviewer oxjR (part 1)

We thank Reviewer oxjR for the focused questions about local data availability, federated aggregation, and candidate budget.

W1. Local data availability and federated aggregation. At 20 local tasks, Table 2's Shared Only row is close to the independent baselines (78.64 on AF and 67.96 on GD). A7 provides the corresponding proxy-only GD comparison: Local 67.64, Shared 66.99, and Personalized 66.34. Table R1 adds matched controls for federated aggregation.

The shared–local decomposition gives a direct explanation. When local data is insufficient, a client can reuse routing structure learned by the other clients. When local data is sufficient, local adaptation can recover more of that structure, while averaging can dilute client-specific preferences. Opposite edge logits cancel in the mean in Eq. (2), and β in Eq. (6) controls the shared–local interpolation. A7 covers 5/10/20 local tasks on AF and 20 tasks on GD; R1 adds the 3-task measurement.

R1 follows the protocol specified in Sections 4.2 and 4.5 and Appendix B.2 of the paper. Predictor-enabled configurations use 60 proxy rounds and rank eight candidates (K=8); predictor-free configurations do not use proxy optimization and deploy one sampled graph (K=1). Across the three clients, the total real-feedback budget is 960 executions for 20 tasks per client and 144 executions for 3 tasks per client (960/20×3), using the same four-round, four-executions-per-task schedule. Each client learns its own edge tensor. Shared configurations average client tensors; configurations without federated aggregation keep them local.

The matched comparisons are:

  • Local only vs. Shared only: identical predictor-free training and deployment; the difference is federated aggregation.
  • Local + Predictor (without federated aggregation) vs. Complete: same pretrained predictor checkpoint, predictor architecture, local-update rule, β=0.9, proxy/feedback schedule, and the same 309 held-out evaluation items per suite. Predictors remain local and are never federated; the comparison changes whether topology updates are averaged across clients.

The newly added 20-task rows are Local only and Local + Predictor; the other 20-task rows reuse the corresponding submitted runs.

Table R1. Matched local-versus-shared results (accuracy %, averaged over the three clients).

ConfigurationAF 3AF 20GD 3GD 20
Local only74.7679.6155.0267.64
Local + Predictor (without federated aggregation)77.0280.9156.3170.55
Shared only (β=0)77.6778.6456.9667.96
Shared + Predictor78.6482.2058.9071.84
Complete (ours)81.2384.4761.1773.14

Paired over the same 309 evaluation items within each suite, using two-sided exact McNemar tests:

ContrastTasksΔ (pp)95% CIbcp
Shared only − Local only (no predictor)3+2.91[0.00, +6.15]1670.093
Shared only − Local only (no predictor, GD)3+1.94[0.00, +3.88]820.109
Shared only − Local only (no predictor)20−0.97[−4.21, +2.27]12150.701
Shared only − Local only (no predictor, GD)20+0.32[−2.59, +3.24]11101.000
Complete − Local + Predictor (without federated aggregation)3+4.21[+0.97, +7.45]1960.014
Complete − Local + Predictor (without federated aggregation, GD)3+4.86[+1.29, +8.41]2490.014
Complete − Local + Predictor (without federated aggregation)20+3.56[+0.65, +6.47]1650.027
Complete − Local + Predictor (without federated aggregation, GD)20+2.59[+0.65, +4.85]1020.039

The matched comparisons show a change in component contribution. When local data is insufficient (3 tasks), federated aggregation has a positive point estimate on both suites. When local data is sufficient (20 tasks), sharing alone is within ±1 point, while the complete configuration remains 3.56 AF / 2.59 GD points above the local predictor baseline. The reported tests isolate aggregation without predictor guidance and compare the complete configuration with its no-federation counterpart; the complete-versus-local comparison includes personalization.

The complete configuration combines federated aggregation with personalization. Its margin over Local + Predictor is 4.2–4.9 points at 3 tasks and 2.6–3.6 points at 20 tasks.

AuthorsOpenReview ↗

Response to Reviewer oxjR (part 2)

W2. Candidate budget. Fig. 4(b) uses the 103 held-out evaluation items of the GD Reason-majority client, with the same executor and judge as the main experiments. It reuses the corresponding learned policies and predictor and varies only the deployment candidate budget K. Thus, it is a single-client deployment diagnostic corresponding to one client row in Table 2, whereas Table 2 reports the fixed-budget results averaged over all three clients. To examine client dependence in candidate scoring, we added a separate deployment-only sweep on all three AF clients. Here, “personalized” denotes the Shared + Personalized + Predictor policy and “shared” denotes Shared + Predictor. We reused the policy and predictor states from Table 2, froze them, and varied only K ∈ {1, 8, 64, 512}; no parameters were retrained and no new MAS executions were collected. Each AF client uses the same held-out split of 103 items as in the main experiments. The table uses the same client set and task coverage as Table 2 and reports a separate frozen-predictor scoring diagnostic.

Clientpers. 1pers. 8pers. 64pers. 512shared 1shared 8shared 64shared 512
AF Code-majority74.1884.6287.4088.1668.0775.9081.2482.61
AF Math-majority80.4291.8793.6194.2875.3682.7787.9588.94
AF Reason-majority71.0582.2985.0386.1166.3174.6880.0281.29
AF macro mean75.2286.2688.6889.5269.9177.7883.0784.28

Each cell reports predictor-estimated accuracy (%) over the held-out items after sampling K candidates for each item and selecting the candidate with the highest Predictor score. The pers. columns use the personalized policy and the shared columns use the shared policy; the macro mean is the unweighted mean of the three AF clients. For example, pers. 8 versus shared 512 compares whether personalization reaches a higher predictor-estimated accuracy with eight candidates than the shared policy reaches with 512 candidates, with the learned policies held fixed.

At K = 8, the personalized policy is above the shared policy on all three AF clients (+8.72, +9.10, and +7.61 points), with a +8.48-point macro-mean difference. At K = 512, the personalized policy at K = 8 remains above the shared policy at K = 512 on all three clients (+2.01, +2.93, and +1.00 points; +1.98 points in macro mean). Thus, the candidate-budget advantage is observed across the three AF clients in this deployment diagnostic, although its magnitude varies by client.

Q1. Technical contribution and validating comparison. The contribution is the shared/personalized decomposition of aligned edge distributions. Predictor guidance reduces the cost of searching those distributions. The matched Shared only versus Local only comparison isolates and measures the aggregation component; Complete versus Local + Predictor without federated aggregation evaluates the full configuration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.