acceptodds
Withdrawn from ICLR 2027

Federated Multi-Agent Orchestration via Consensus-guided Interaction Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In large language model (LLM)-based multi-agent systems (MAS), the orchestration of inter-agent communication plays a crucial role in improving reasoning quality and collaboration efficiency. However, existing methods typically either learn a dedicated orchestrator for each domain or train a single orchestrator on centrally pooled multi-domain data. These paradigms limit knowledge integration across heterogeneous clients with siloed data, hindering both the learning of a unified orchestrator that accommodates diverse agent configurations and generalization to unseen clients. To address these limitations, we propose FedGLV, a federated MAS collaboration framework that learns transferable interaction dynamics for heterogeneous multi-agent orchestration. Specifically, to characterize heterogeneous inter-agent interactions across clients within a common dynamical formulation, we introduce an amortized relational dynamics model inspired by generalized Lotka–Volterra dynamics. Conditioned on queries and agent roles, this model couples latent relational-state evolution with adaptive topology generation, supporting varying numbers of agents and communication steps. At the server, data-free structural consensus distillation is introduced to integrate orchestration knowledge across heterogeneous clients. Concretely, guided by leave-one-out (LOO) valuation, FedGLV aggregates client soft topologies obtained using the server-generated probe bank into a dynamic structural consensus. This consensus provides an explicit structural supervision signal for the global orchestrator to learn transferable interaction patterns. Experiments on six source datasets for federated training and three unseen-client benchmarks demonstrate the effectiveness and generalizability of FedGLV.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_naZ3OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets to provide a method for different clients, with its own agent system working on different private task and dataset, to train a agent communication controller together via federated learning, to further improve the cooperation between agents and get a better task performance. The scope of this paper is to optimize the controller with LLM or agent untrained. During each round, on the client side, the GLV-inspired controller is used to model the hidden state evolution for generating the communication topology of agents, and the controller then is updated by the final answer for the local task. Then the server will gather all the controller weights from local clients and carry out the weight aggregation and the knowledge distillation by probe and LOO to update the server-side controller. On the evaluation, 6 clients were evaluated on several LLM benchmarks includes Q&A, math and HumanEval and so on. What's more, ablation study on different components as well as unseen task generalization were also evaluated.

Critical strengths
  1. This paper designed a unified agent communication controller for even heterogenous clients: (1) the controller can handle different tasks, agent configuration as well as the communication topologies and steps. (2) It did provide a complete Federated learning method to train the controller for each clients without sharing the private dataset and agent configuration and so on.

  2. Evaluation including ablation on the server side as well as the unseen task generalization provided more evidence: (1) ablation study for different components including GLV, probe, LOO and so on. (2) On unseen task, proposed method has more uplift compared to the baseline, which provided important evidence.

Critical weaknesses
  1. Currently the cross client-sharing training benefit was not directly demonstrated: (1 ) there should be a ablation study on GLV controller training result on solo client compared with federated GLV controller training, under the same local data, local update as well as the agent configuration and so on. (2) Currently the main results include the ablation on different controller model and training method, but it didn't isolate the contribution of federated learning. (3) Since the federated collaborative learning is the core part of this paper, lacking of the demonstration of benefit of federated learning reduced the technical soundness.

  2. Currently the ablation study can not demonstrate the necessary of GLV:(1) current result can only demonstrate the benefit of dynamic state evolution by comparing to other baselines, however, it lacks of comparation with other simple dynamic controller such as simple RNN controller. (2) This is important because the GLV is the core design of the FedGLV. Lack of ablation study on other dynamic baselines also reduce the soundness.

Questions for authors

My questions are closely related to the critical weakness: author should provide (1) more evidence of the isolated federated collaborated learning benefit and (2) more evidence of the necessary of GLV instead of other simple dynamic controller algorithm such as simple RNN , MLP with step embedding and so on.

Reviewer_BLS4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedGLV, a federated multi-agent orchestration framework for heterogeneous clients. It introduces a relational-state model inspired by generalized Lotka-Volterra dynamics to generate communication topologies that vary with the query, agent roles, and communication step. Each client optimizes its orchestrator using task-correctness rewards on private data. On a shared, answer-free probe bank, the server computes leave-one-out (LOO) weights from normalized relational-state agreement, forms a weighted consensus of client soft topologies, and distills this structural consensus into a global orchestrator. The experiments cover six source clients used for federated training and three held-out clients, with additional analyses of heterogeneous agent configurations, sparse communication, dynamic relational states, and server-side aggregation.

Critical strengths

Transfer to unseen clients: The final global orchestrator is deployed to three held-out clients without target-domain parameter updates and achieves the best average result. The server-side ablations also show that the full method outperforms FedAvg and uniform structural distillation on average.

Critical weaknesses
  1. The necessity of cross-task orchestration sharing is not directly validated: Each client has its own tasks, role configuration, and correctness feedback, and could in principle learn a task-specific orchestrator independently. The paper does not show that receiving the server-fused orchestrator and then adapting it locally outperforms local training without subsequent global broadcasts. It therefore remains unclear whether cross-domain sharing consistently benefits the source clients or may cause negative transfer. The w/ FedAvg and w/o LOO variants still broadcast global models and do not replace a Local-only FedGLV baseline with the same architecture and training budget.

  2. Answer-free state agreement is not equivalent to task utility: The server assigns weights using relational-state agreement on an answer-free probe bank, but this signal cannot by itself determine whether a communication structure improves task accuracy. High agreement may only indicate that several clients behave similarly rather than effectively, while a useful but distinctive client may be downweighted because it disagrees with the others. Although Table 3 shows an average improvement over uniform weighting, the paper does not validate the relationship between LOO weights and actual validation performance or client-level marginal utility.

  3. A key LOO-weighted parameter-aggregation ablation is missing: The paper compares standard FedAvg, uniform topology distillation, and LOO-weighted topology distillation, but does not test direct parameter aggregation using client-level weights derived from the LOO scores. The existing results partially isolate the effects of structural distillation and LOO weighting, but they do not establish whether the benefit of LOO requires topology-space distillation or could also be obtained through generic client reweighting.

  4. The scales of x and b and the stability of the exponential update are insufficiently analyzed: The initial state x and growth rate b are produced by two separate Softplus heads, which ensure positivity but impose no explicit normalization, upper bound, or cross-head scale calibration. These quantities are combined with the interaction term inside an exponential update, which could in principle produce overly large growth or decay. Although the saturating interaction function, self-regulation term, and step size provide indirect control, the paper does not report the empirical distributions of these variables or analyze the stability of the discrete dynamics. Moreover, LOO agreement is computed from normalized state vectors, whereas the dynamics use the unnormalized state magnitudes and topology generation is modulated by kappa(x); normalization may therefore hide cross-client scale differences that still affect the learned behavior.

Questions for authors
  1. Does Table 1 evaluate the final global orchestrator or the locally adapted orchestrator for each source client? Could the authors add a Local-only FedGLV baseline with the same architecture and training budget, and report per-client performance before global broadcast, immediately after broadcast, and after local adaptation?

  2. Could the authors report the correlation between LOO weights and client validation performance or actual downstream marginal contributions? A small labeled probe set could also provide a diagnostic oracle comparison without being part of the deployed method.

  3. Could the authors add LOO-weighted parameter aggregation to complete the factorial comparison between parameter-space versus topology-space aggregation and uniform versus LOO-based weighting?

  4. Could the authors report the distributions of x0x^0, bb, xsx^s, and the exponent increment across clients, agent counts, and communication steps, and state whether numerical overflow, gradient explosion, or state collapse was observed?

Reviewer_m2WUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

FedGLV trains one generator of agent communication graphs across federated clients from different domains. Each agent carries a scalar state, evolved through a Lotka–Volterra-style update, that gates which edges form. The server runs each client's model on 40 probe queries, weights clients by how muuch they agree with each other, and distills the rsult into a global model. Using GPT-4o-mini with six source datasets of 20 training queries each , it beats the best baseline by 0.86 points on source clients and by about 2 points on unseen ones.

Critical strengths

Replacing distillation with plain FedAvg costs about 2 points on unseen clients and loses on all trhee, which is the cleanest result in the paper. The federated setting is new and well defined.

Critical weaknesses

The per-dataset gains of under one point sit inside the repoted error bars. The step counts match the best values from sweeeps run on test data, and one key weight was tuned on an unseen client's test set. FedGLV also gets hand-written specialist roles for the unseen domains. No control with the same roles and fixed or random graphs are reported, so nothing ties the gains to the method.

Questions for authors

How many runs are behind each number? Did the baselines use the same roles? Was there a validation split? For futture work, the agent states could react to actual messages, since they are currently computed before anyone speaks.

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

Related papers

Discussion (0)

Sign in to comment.