acceptodds
Under review as a conference paper at ICLR 2027

Learning Who Talks to Whom: Role-Aware Communication Topology Scheduling for Efficient LLM Multi-Agent Collaboration

Abstract

Multi-agent systems driven by large language models (LLMs) rely on structured communication topologies to facilitate information exchange and collaborative reasoning. However, effectively scheduling these topologies to dynamically determine who talks to whom and when to communicate remains a fundamental challenge in balancing collective reasoning performance against exponential token overhead. Existing topology scheduling approaches either enforce static communication graphs that cannot adapt to dynamic problem-solving states, or overlook agent role complementarity while failing to explicitly incorporate the synthesized graph structure into team-level decision evaluation. To tackle these limitations, we propose ACTS, a role-aware adaptive communication topology scheduling framework for efficient LLM multi-agent collaboration. Orchestrated within a Centralized Training with Decentralized Execution (CTDE) paradigm, ACTS introduces a novel dual-attention mechanism. Specifically, a Role-conditioned Relational Attention module dynamically selects variable-cardinality interaction partners based on agent roles and evolving execution states, while a Topological Attention Mixer explicitly embeds the directed communication topology into team value decomposition for precise credit assignment. Comprehensive evaluations across 9 challenging benchmarks covering code generation, multi-step reasoning, general knowledge, and competitive mathematics demonstrate that ACTS consistently outperforms state-of-the-art baselines, boosting average accuracy by up to 6.19 percentage points while reducing token consumption by over 15%. Code is available at https://anonymous.4open.science/r/acts-B1F2/.

Primary area:
reinforcement learning
Keywords:
LLM Multi-Agent Systems, Cooperative Multi-Agent Reinforcement Learning, Value Function Factorization, Topology Scheduling, Multi-Head Attention, Centralized Training with Decentralized Execution

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_rz8JOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the accuracy–cost trade-off in LLM-based multi-agent collaboration by learning when and with whom agents should communicate. It proposes ACTS, a communication topology scheduling framework that keeps the underlying LLMs frozen and follows centralized training with decentralized execution. Local controllers select communication modes, while role-aware relational attention guides partner selection to form adaptive directed communication graphs. A topology-aware value mixing mechanism enables joint training of these decisions under an objective balancing task accuracy and token cost. Experiments across nine benchmarks with two LLM backbones report improved accuracy and communication efficiency over the evaluated multi-agent baselines.

Critical strengths

I appreciate the paper’s contribution toward learning adaptive communication topologies for LLM-based multi-agent collaboration. The main text and appendix together provide a largely comprehensive and clear description of the method. Most methodological details, including topology construction, value mixing, and the training and execution procedures, are explained explicitly, making the overall framework easy to follow and assess.

Critical weaknesses

1.The evaluation is restricted to three-agent teams, providing limited structural complexity for a method centered on communication topology optimization. With at most two candidate partners per agent, this setting offers limited scope to assess partner selection and the claimed benefits of variable-cardinality routing. Experiments with larger teams are needed to establish whether the accuracy and efficiency gains persist as the communication structure becomes more complex.

2.The authors describe relational attention as selecting communication partners based on semantic relevance. However, learned attention scores alone do not establish this interpretation, and the observation specification in Appendix E.2 does not clarify how message content enters the controller. The authors should explain how generated and received messages are encoded into hith_i^t to support semantic relevance-based partner selection.

3.Topology-aware value mixing is a key contribution of the paper. However, the proposed module differs from QMIX in both its use of explicit topology information and its attention-based mixing architecture. Replacing the entire module with QMIX does not isolate these contributions, making it unclear whether the gains stem from topology information or the architectural change. An additional ablation retaining the attention-based mixer while removing explicit topology features would clarify this distinction.

Questions for authors

1.Although ACTS reduces token consumption, according to Equations (4)–(5) relational attention requires each agent to access representations from all other agents. Does this introduce additional communication and computational overhead that is not captured by the token-based cost metric? How does this overhead scale with the number of agents, and how might it affect the overall efficiency gains in larger teams?

2.The motivation contrasts ACTS with centralized methods that generate a topology before execution. However, I think a centralized controller could also update the topology at each communication round. If the same agent information is available to a centralized controller, what motivates the choice of decentralized decision-making in ACTS, and what advantages or trade-offs do the authors see relative to centralized control?

4 comments
AuthorsOpenReview ↗

Response to Reviewer rz8J — Part 1/4: Larger teams and partner selection (W1)

Dear Reviewer rz8J,

Thank you for recognizing the contribution of learned communication topologies and for identifying the limitation of a three-agent test. With only two possible peers, partner selection can resemble mode selection. A larger team provides a more informative test of whether learned routing continues to improve coordination.

From three agents to larger teams. In the main three-agent comparison, ACTS gains 3.17/6.19 pp in macro accuracy with 19.43%/18.75% fewer inference tokens than Agent Q-Mix on GPT-OSS-120B/DeepSeek-V4-Flash (Tables 1 and 3). The larger-team evaluation tests whether this joint advantage persists with more candidate communication partners.

Larger-team evaluation. Section 4.3.4 and Appendix E.8 evaluate N = 6 on all nine benchmarks with both backbones, increasing each agent's potential partners from two to five. ACTS and Agent Q-Mix use the same 60-episode training budget and at most five communication rounds.

One mode, different partner sets. A six-agent sender still makes the same six-way mode decision. Under attn-broadcast, it requests a link to each active peer whose attention to the sender exceeds the threshold. Changing these scores changes both the identities and the number of recipients while the mode stays fixed (Section 3.2). This separates learned partner selection from mode selection.

The table combines the main N = 3 comparison with the added N = 6 evaluation (Tables 1 and 3). ACC is nine-benchmark macro accuracy (%). Tokens are thousands per question.

BackboneNACTS ACCAgent Q-Mix ACCACTS tokensAgent Q-Mix tokens
GPT-OSS-120B375.0171.835.977.41
GPT-OSS-120B675.4171.009.9012.72
DeepSeek-V4-Flash369.4363.244.034.96
DeepSeek-V4-Flash670.1964.287.099.32

The larger-team evaluation yields two consistent improvements:

  • Higher accuracy: ACTS gains 4.40 and 5.91 percentage points on GPT-OSS and DeepSeek, computed from benchmark-level counts before rounding.
  • Lower inference cost: token use falls by 22.17% and 23.93%, respectively.

Participation and connectivity. Table 10 reports 4.11/4.00 active agents and 5.80/5.58 directed edges per executed round for ACTS on GPT-OSS/DeepSeek, compared with 5.19/5.10 agents and 9.69/9.40 edges for Agent Q-Mix. Higher accuracy accompanies more selective participation and routing within the larger team. Edge counts describe connectivity, while the fixed-partner ablation in Table 2(b) tests partner choice more directly: fixing partners lowers accuracy and raises token use on both backbones. Together, these results connect the larger-team advantage to the purpose of learning whom agents should communicate with.

The N = 6 results use benchmark-level correct/total counts. The three-training-run statistics in the manuscript describe the native N = 3 setting. Expanding the team increases token use for both methods. Within the larger-team setting, ACTS retains its accuracy and token advantage over Agent Q-Mix.

Revision locations: Section 4.3.4, Table 3, and Appendix E.8 (Table 10). The corresponding results and discussion are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you again for encouraging a more informative test of the scheduler's routing choices.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer rz8J — Part 2/4: Shared execution and computational overhead (Q1)

Thank you for asking whether token savings translate into practical efficiency once shared attention is included. The analysis distinguishes execution requirements, measured inference latency, and training costs.

Execution requirements. Sections 3.2–3.4 and Appendix D describe a shared orchestrator that computes GNN–GRU control representations and relational attention across agents. Attention scores candidate partners and refines the features used for per-agent mode decisions. Modes and partner scores jointly determine participation and directed communication. The topology-conditioned mixer supplies the cooperative objective during training. The execution network combines the learned agent policy with shared relational attention.

Inference overhead. Pairwise attention scoring costs O(N² d_rel) per round. Appendix E.5.2 (Table 8) reports accumulated per-question latency: controller time (Ctrl.) in ms/question and end-to-end time (E2E) in s/question.

BackboneNCtrl. ACTSCtrl. Agent Q-MixE2E ACTSE2E Agent Q-Mix
GPT-OSS-120B34.75.818.621.9
GPT-OSS-120B68.411.831.739.8
DeepSeek-V4-Flash34.55.616.820.1
DeepSeek-V4-Flash68.111.529.436.2

Controller timing covers observation construction, the GNN–GRU policy and relational attention, action selection, and graph construction. End-to-end timing additionally includes the agent and final-decision LLM calls. For six-agent GPT-OSS, these measurements are 8.4 ms and 31.7 s per question, respectively, versus 39.8 s end-to-end for Agent Q-Mix. ACTS records lower controller and end-to-end latency in all four configurations. Thus, the token advantage is accompanied by lower total inference time with shared computation included.

The measurements describe shared-orchestrator execution. O(N² d_rel) covers pairwise attention. A deployment across separate nodes would additionally incur representation-transfer costs.

Training costs. Table 8 reports input + output training tokens (M = millions), rollout time, and optimization time (h = hours). GPT-OSS and DeepSeek abbreviate the backbones above.

BackboneNMethodTrain tokens (M)Rollout (h)Optim. (h)
GPT-OSS3ACTS0.384.620.18
GPT-OSS3Agent Q-Mix0.475.470.15
GPT-OSS6ACTS0.647.910.31
GPT-OSS6Agent Q-Mix0.839.860.27
DeepSeek3ACTS0.334.180.17
DeepSeek3Agent Q-Mix0.414.960.14
DeepSeek6ACTS0.587.150.29
DeepSeek6Agent Q-Mix0.748.920.25

ACTS uses fewer training-interaction tokens and less rollout time in every configuration, with slightly higher optimization time. Together with the inference measurements, these results show that the learned scheduler's interaction savings extend to both training and execution.

Revision locations: Sections 3.2–3.4, Appendix D, and Appendix E.5.2 (Table 8). The execution clarification and cost measurements are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you again for helping us connect token efficiency to the computation required in practice.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer rz8J — Part 3/4: Message semantics and factorized control (W2, Q2)

Thank you for identifying the ambiguity in semantic partner selection and for distinguishing dynamic topology updates from centralized control. We have clarified both points and connected the architectural choices to controlled experiments.

Scheduler inputs and identity. Observations include a deterministic 32-bit question hash, a 16-dimensional one-hot agent identifier as used in Agent Q-Mix’s public implementation, answer availability, degree information, normalized cumulative token use, and previous activity flags (Sections 3.1–3.2, Appendix E.2). The graph separately enters the GNN, and the GRU summarizes graph-conditioned control history. Identity information has distinct uses:

  • Observation identity distinguishes agents with fixed role assignments.
  • Role prompts condition LLM calls.
  • Learned agent-index embeddings condition the training-only mixer.

Generated and received answers supply the LLM language contexts after routing. The shared orchestrator computes each agent's recurrent representation from task and identity features, execution statistics, and the current graph. Pairwise attention scores these control states, expressing control-state-conditioned partner preferences, rather than semantic agreement or relevance between answer texts. Figure 4 (p.6) makes the distinction visible: training is on the left, scheduling in the middle, and answer generation and aggregation on the right. Scheduler parameters stay fixed at inference while control states and graphs evolve.

Controlled evidence. Table 2 evaluates the following interventions within ACTS. ACC is nine-benchmark macro accuracy (%), and tokens are thousands per question.

ConfigurationGPT-OSS ACCTokensDeepSeek ACCTokens
Full ACTS75.015.9769.434.03
Without observation identity73.406.8067.104.60
Fixed communication partners72.508.1066.106.20
ACTS mixer → QMIX71.837.4163.244.96
  • Agent identity: removing observation identity retains role prompts and mixer embeddings, isolating this input's contribution.
  • Partner selection: fixing partners lowers accuracy and increases tokens on both backbones.
  • Team-value learning: replacing only the ACTS mixer with QMIX also worsens both metrics. This controlled ablation retains the ACTS execution network and changes its training mixer.

Rationale for factorized control. We agree that a centralized controller can update its graph every round. ACTS factorizes mode decisions so that each agent selects its highest-valued mode from six scores, while training uses one cooperative team objective. Following QMIX, nonnegative mixing weights ensure that these greedy choices jointly maximize the team value represented by the mixer at fixed pre-decision context (Eq. (12), Appendix C.3). This provides greedy consistency within the represented monotonic value class. Execution still uses shared control representations for relational attention. The controlled comparisons above assess the contributions of identity, partner selection, and value mixing within this architecture.

Revision locations: Sections 2.1, 3.1–3.4 and 4.3.3, Figure 4, Table 2, and Appendices C.3, D, E.2 and E.7. The corresponding clarifications and ablation discussion are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you again for helping us clarify how learned topology scheduling supports LLM collaboration.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer rz8J — Part 4/4: Isolating explicit neighborhood inputs (W3)

Thank you for identifying that replacing the mixer with QMIX changes both its architecture and its topology inputs. The targeted ablation follows your suggestion to retain the attention architecture. Building on attention-based value mixing as studied in Qatten, it holds the mixer architecture fixed and varies only the directed-neighborhood channels, directly testing the topology-conditioning mechanism.

Targeted intervention. The new variant in Table 2(d) and Appendix E.7 changes only the mixer's explicit neighborhood inputs:

  • Zeroed inputs: the incoming- and outgoing-neighborhood mean aggregates in Eq. (9).
  • Retained components: the attention-mixer architecture, learned agent-index embeddings, and execution-network architecture.
  • Evaluation procedure: retrain the variant from scratch and evaluate the resulting scheduler.

ACC is nine-benchmark macro accuracy (%). Tokens are thousands per question.

Mixer configurationGPT-OSS ACC (%)GPT-OSS tokensDeepSeek ACC (%)DeepSeek tokens
Full ACTS75.015.9769.434.03
Zero incoming/outgoing neighborhood means72.677.4067.116.72
Change vs. full ACTS−2.34 pp+1.43−2.32 pp+2.69

Zeroing the neighborhood inputs reduces accuracy by 2.34/2.32 percentage points and increases token use by 1.43/2.69 thousand tokens per question on GPT-OSS/DeepSeek. The consistent direction across backbones supports the contribution of explicit directed-neighborhood conditioning within the same attention architecture. These inputs shape training, and their benefit appears in the accuracy and communication cost of the learned execution policy.

These two channels separately summarize the peers connected to an agent in each direction on the pre-decision graph. Graph propagation and relational partner selection remain available at execution. The intervention therefore isolates these explicit training-time neighborhood inputs within the same attention architecture, connecting topology-conditioned value learning to the resulting policy's coordination performance.

Reporting basis. Table 9 provides benchmark-level correct/total counts for the added variant. Its accuracy is computed from these counts, while full ACTS uses the mean of three independent scheduler-training runs as the reference.

Revision locations: Section 4.3.3, Table 2(a,d), and Appendix E.7 (Table 9). The intervention, results, and discussion are highlighted in blue\textcolor{blue}{\text{in blue}}.

We appreciate this suggestion, which enabled a more direct test of why topology enters the training objective.

Best regards,
The authors

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

This paper proposes ACTS, a role-aware scheduler that learns when and between which LLM agents communication should occur. It combines relational attention for partner selection with topology-aware value mixing, optimizing task accuracy and token efficiency while keeping the LLM backbones frozen. Experiments on nine benchmarks and two backbones report improvements over existing coordination methods, with ablations examining the main components.

Critical strengths
  1. The studied problem is well motivated and valuable. Adapting communication to agent roles and task progress is a natural approach to reducing redundant exchanges while preserving useful collaboration.
  2. The method combines role-aware partner selection with topology-aware value mixing in a coherent framework. The nonnegative mixer preserves greedy-action consistency under the stated fixed-context assumption.
  3. The experiments cover nine benchmarks, four task domains, and two frozen LLM backbones. Comparisons with multiple coordination methods and component ablations provide useful initial evidence for the proposed design.
Critical weaknesses
  1. The methodology section is quite difficult to read and follow, and some central implementation details remain unclear even after consulting the appendices. For example, the meanings and consequences of aoffa^{off} and asoloa^{solo} in Eq. (1) are not explained when introduced; the contents and encoding of the local observation oito_i^t in Eq. (3) are unclear; and the definitions of broadcast/attn-broadcast in Eq. (6) are very confusing. These essential concepts, together with the roles of the two attention modules and the overall training and execution procedure, should be explained clearly in the main paper rather than scattered across the appendix. The authors are strongly suggested to prioritize these explanations over some of the figures in the main paper. Furthermore, Appendix E.2 describes an observation containing a task embedding and bookkeeping features, without explaining how generated answers or received messages are encoded, whereas Algorithm 2 describes observations built from interaction histories and delivered messages. Eqs. (4) and (5) also use other agents' hidden states during action selection, leaving unclear how these states are obtained under the stated local-observation restriction. These gaps affect both reproducibility and the assessment of the claimed adaptation to evolving execution context.

  2. The discussion and experimental comparison with methods that adapt communication structures during execution are largely limited to Agent Q-Mix. However, several other relevant approaches exist. DyLAN, already included in the references, adjusts subsequent agent participation based on current responses; DyTopo reconstructs directed communication graphs each round through semantic matching; and TodyComm learns round-wise communication topologies through policy gradients. Although their mechanisms differ from ACTS, they address closely related dynamic coordination problems. The authors are suggested to discuss these approaches and clarify the methodological differences, and to include at least one additional representative method in the experiments under comparable backbones and budgets.

  3. The efficiency advantage is not sufficiently established when training cost is considered. They reported 50 episodes training steps, but the number of episodes alone does not establish computational efficiency or exact training cost. The paper reports inference-token savings without comparing the actual training time, rollout token consumption, or optimization cost against other methods. Consequently, it is difficult to judge ACTS's (which incorporates several different modules together including GNN, GRU, Attention, Q-nework) overall accuracy-cost advantage over alternatives with lower training overhead or no training requirement.

  4. The reliability and generality of the reported improvements require stronger evidence. Several evaluation sets are small: MMLU-Pro contains 70 questions, while each AIME benchmark and HMMT contain 30. Although results are averaged over three seeds, no variability estimates or paired significance analyses are reported. The source of the 15 training examples per domain and their separation from evaluation data are also unclear. In addition, the role ablation needs clarification: Section 4.3.3 retains role-specific prompts when removing role embeddings, whereas Appendix E.6.5 describes a role-agnostic prompt ablation without explaining its relationship to Table 2. This makes it difficult to identify which source of role information accounts for the reported gains. Finally, all experiments use only three agents and two or three communication rounds, leaving scalability and longer-horizon coordination untested.

Questions for authors

Most of my questions are already discussed in the Weaknesses section. I briefly restate some of them here for clarity:

  1. How are observations encoded, local Q-values computed, and other agents’ hidden states accessed during decentralized execution?
  2. Can you add another dynamic communication baseline and compare total training and inference costs under matched settings?
  3. Can you clarify the training/evaluation split and report uncertainty estimates for the main improvements?
7 comments
AuthorsOpenReview ↗

Response to Reviewer ohrD — Part 1/5: Observations and communication actions (W1, Q1)

Dear Reviewer ohrD,

We appreciate your careful examination of the implementation and your helpful feedback. We agree that the paper should specify what the controller observes and how its decisions become LLM communication. ACTS learns agent participation and directed communication at each round to support effective collaboration with fewer LLM tokens.

From a control decision to an answer. Consider a round in which A requests an answer from B. The scheduler decides whether to request the link using the agents’ control states. Once the graph is constructed, B’s available answer text enters A’s LLM context, together with the question and role prompt. Scheduling and answer generation therefore use different information (Figure 4).

Scheduler inputs. Sections 3.1–3.2 and Appendix E.2 now distinguish three input groups:

  • Task and agent identity: we use a deterministic 32-bit question hash expanded into binary features, together with a 16-dimensional one-hot agent identifier following Agent Q-Mix’s public implementation.
  • Execution statistics: answer availability, the sum of incoming and outgoing degrees, normalized and clipped cumulative team-token usage, and previous activity indicators.
  • Topology: the directed adjacency matrix, supplied separately to the GNN.

The identifier distinguishes agents under fixed role assignments, enabling agent-specific scheduling. Role prompts condition the LLMs, while separate learned agent-index embeddings condition the training mixer. The hash represents the task through a deterministic binary code, and the availability flag indicates whether an answer exists. Neither encodes answer content. The GNN and GRU instead summarize graph-conditioned control features over time, tracking participation, connectivity, and expenditure. Algorithm 2 follows this same input specification.

Action semantics. Section 3.1 and Equations (6)–(7) distinguish LLM invocation from link requests:

  • off: skip the current LLM call and remove incident links for that round. The agent can return in a subsequent non-terminal round.
  • solo: invoke the LLM without requesting links. Other active agents may still request links involving it.
  • broadcast: request outgoing links to all other active agents.
  • receive: request incoming links from all other active agents.
  • attn-receive: A selects active senders whose scores in A’s attention row exceed that row’s mean.
  • attn-broadcast: A selects active recipients using its attention column, which records how strongly each recipient attends to A.

For both attention modes, selection uses scores above the corresponding row or column mean over all N entries, including the zero diagonal. Either endpoint can request an edge. Requests are combined by union, excluding links incident on off agents. Appendix B provides a numerical example. Thus, A can retain the same attn-receive mode while changing whose answers reach it. Modes control participation and exchange type, while attention adapts the number and identity of communication partners.

Revision locations: Sections 3.1–3.2, Algorithm 2, and Appendices B and E.2–E.3. The corresponding clarifications are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you for helping us make these implementation choices explicit in the presentation.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer ohrD — Part 2/5: Shared execution and training-only value mixing (W1, Q1)

Your question about access to peer states identifies an important distinction between factorized decisions and shared computation. The revised execution model and Figure 4 specify the roles of both attention modules. A controlled ablation isolates the mixer's topology inputs.

What is available when A chooses B? In Equations (3)–(5), the GNN propagates observations over the current graph and the GRU updates recurrent control states. A shared orchestrator computes these states for the team and then applies relational attention to them. This gives A’s decision access to B’s control representation at inference. Relational attention scores directed partners and refines action features. A shared Q-head outputs six mode values per agent for greedy selection. Execution therefore combines factorized action selection with shared control representations.

Training–execution interface. Figure 4 presents the complete workflow: the left panel trains team values, the middle panel selects modes and partners, and the right panel routes answers, invokes active LLMs, and aggregates their outputs. The scheduler decides the communication structure, while the LLMs generate the task answers:

ComponentTraining functionInference function
GNN–GRU + Q-headLearn control representations and mode valuesCompute per-agent mode values
Relational attentionLearn partner scores and shared action featuresUpdate scores and features each round
Topological attention mixerLearn team values through TD lossTraining only

At inference, scheduler parameters remain fixed, while control states, actions, and graphs update every round. Relational attention continues to select partners. The global-state mixer supplies the team-value objective during training. Deployment uses the learned GNN–GRU, Q-head, and relational attention to update communication decisions.

Evidence for topology-conditioned learning. Table 2(d) zeros only the mixer's incoming/outgoing neighborhood means and retrains with the remaining mixer and execution architectures unchanged. ACC is macro accuracy (%) across nine benchmarks. Tokens are thousands per question. The ablation uses benchmark-level counts, compared with the full-model three-run reference.

Mixer inputsGPT-OSS ACCTokensDeepSeek ACCTokens
Full ACTS75.015.9769.434.03
Neighborhood means zeroed72.677.4067.116.72

Removing these inputs reduces accuracy by 2.34 and 2.32 percentage points and increases token use on both backbones. With the attention architecture retained, this experiment supports the contribution of directed-neighborhood information to learning accurate, token-efficient communication schedules.

Revision locations: Sections 3.2–3.4, Figure 4, Appendix D, and Section 4.3.3 / Table 2(d), with the intervention detailed in Appendix E.7. The corresponding text and results are highlighted in blue\textcolor{blue}{\text{in blue}}.

We appreciate the opportunity to clarify how shared execution and training-only value mixing work together.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer ohrD — Part 3/5: Dynamic baselines and computational costs (W2–W3, Q2)

We agree that dynamic communication should be discussed beyond a single architecture and that efficiency should include the cost of learning the scheduler. The revised comparison distinguishes dynamic communication architectures and reports measured training and inference costs under matched settings.

Relationship to dynamic communication methods. Section 2.1 discusses the mechanisms of the suggested approaches:

  • DyLAN adapts participation using response-based ranking.
  • DyTopo reconstructs graphs through semantic matching.
  • TodyComm learns communication structures through policy gradients.

ACTS learns participation and relational partner selection through cooperative value decomposition, with directed-neighborhood information conditioning the training mixer. The empirical comparisons below are the baselines evaluated in Tables 1 and 7, while these three approaches inform the related-work discussion.

Main-paper accuracy–cost comparison. On HotpotQA, Tables 1 and 7 report accuracy (%) and inference tokens (K/question) for representative static, centralized adaptive, and MARL methods under the same backbone, team size, round budget, and evaluation protocol (Section 4.1):

MethodGPT-OSS-120B ACCTokensDeepSeek-V4-Flash ACCTokens
LangGraph40.0016.446.6211.0
G-Designer34.7312.546.628.4
GTD35.9512.147.978.1
Agent Q-Mix40.008.744.005.8
ACTS46.677.150.674.8

ACTS achieves higher accuracy with fewer tokens than these multi-agent baselines on both backbones. GTD generates a task-specific graph, illustrating that this comparison already includes adaptive topology design. Matched-comparison rationale. Agent Q-Mix shares cooperative value decomposition, GNN–GRU representations, and per-agent modes. Matching backbones, task instances, and training budgets provides a direct comparison for ACTS’s relational routing and topology-conditioned value learning.

Measured costs. Both methods use 60 training episodes, at most five communication rounds, and the matched-data protocol in Section 4.1 / Appendix E.1. Appendix E.5.2 / Table 8 reports three- and six-agent costs on both backbones. The three-agent GPT-OSS-120B results are:

Cost componentACTSAgent Q-Mix
Training input + output tokens (M)0.380.47
Training rollout time (hours)4.625.47
Scheduler optimization time (hours)0.180.15
Controller latency (ms/question)4.75.8
End-to-end latency (s/question)18.621.9

For this three-agent GPT-OSS setting, ACTS uses 0.38M versus 0.47M training tokens. The increase in optimization time (0.18 versus 0.15 hours) is smaller than the reduction in rollout time (4.62 versus 5.47 hours). These separately measured components show that savings in training tokens and LLM rollout time already appear relative to Agent Q-Mix. At inference, end-to-end latency falls from 21.9 to 18.6 seconds per question. The comparison concerns two learned schedulers under matched training budgets.

Across the nine-benchmark suite, ACTS uses 19.43%/18.75% fewer inference tokens on GPT-OSS-120B/DeepSeek-V4-Flash (Table 3). Across both backbones and N ∈ {3, 6}, it uses 19.15–22.89% fewer training tokens, with lower rollout and end-to-end times (Table 8). Appendix E.5.2 specifies hardware, warm-up, synchronization, and per-question timing. End-to-end measurements include the shared controller, agent calls, and final aggregation. Thus, the accuracy–token advantage is accompanied by lower measured execution latency.

Revision locations: Section 2.1 and Appendix E.5.2 / Table 8. The added discussion and cost evidence are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you for encouraging a more complete account of the cost of learning and deploying the scheduler.

Best regards,
The authors

Reviewer_G6taOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ACTS, a reinforcement-learning framework for adaptive communication topology scheduling in LLM-based multi-agent systems. It combines role-conditioned relational attention for selecting communication partners with a topological attention mixer for topology-aware value factorization. Experiments on nine benchmarks show improved accuracy and reduced token consumption compared with existing baselines, including Agent Q-Mix.

Critical strengths
  1. The paper addresses a problem in LLM-based multi-agent systems: balancing collaborative reasoning performance against communication cost. The evaluation covers nine benchmarks across multiple task domains and reports improvements in both accuracy and token efficiency.
  2. The paper provides component-level ablations and communication-pattern analyses. These experiments offer some evidence that role conditioning, adaptive routing, and topology-aware value mixing contribute to the reported performance.
Critical weaknesses

Insufficient evidence of generalization. The scheduler is trained on only 15 examples per domain for 50 episodes, raising concerns about overfitting to the training tasks and backbone-specific interaction patterns. Although the evaluation spans multiple benchmarks and two backbones, the paper does not sufficiently establish whether the learned communication policy generalizes to unseen task distributions or transfers across backbones.

Questions for authors
  1. Given the close overlap with Agent Q-Mix, what specific capability does ACTS enable that cannot be achieved by existing communication schedulers with comparable modifications? Can the authors provide evidence beyond component ablations that isolates the contribution of the proposed formulation?
  2. How robust is ACTS to the limited training data? Were ACTS and Agent Q-Mix trained on exactly the same task instances, data splits, and training budgets? Can the authors provide cross-task or cross-backbone transfer experiments to demonstrate that the learned scheduler generalizes beyond its training configuration?
4 comments
AuthorsOpenReview ↗

Response to Reviewer G6ta — Part 1/4: Contribution and motivation (Q1)

Dear Reviewer G6ta,

Thank you for your careful review and for recognizing the importance of balancing collaborative reasoning with communication cost. Your questions about the distinction from Agent Q-Mix and policy generalization helped us sharpen the contribution and its evaluation. The corresponding revisions are highlighted in blue\textcolor{blue}{\text{in blue}}.

What does ACTS add to mode learning? ACTS learns who participates and who communicates with whom through cooperative MARL, balancing answer quality and token cost. Agent Q-Mix changes its graph across rounds through communication modes, with predefined edge-construction rules once those modes are fixed. ACTS combines six modes with learned, directed partner selection.

Same mode, different partners. Consider three active agents A, B, and C, with an answer from A available for routing. A can keep attn-broadcast while B and C keep solo, yet request only B in one round and only C in the next as attention scores change. Here, solo makes an LLM call and requests no links itself, but accepts links requested by another agent. Thus only the selected recipient receives A's answer through the new graph. If both candidates pass the threshold, A requests both. This permits different recipients and group sizes within the same mode (Section 3.2, Appendix B).

How does the team learn when communication is worthwhile? Relational attention both scores partners and refines the features used to value each mode. Team-level TD learning therefore trains the scoring network through the action-value features, while thresholding constructs the discrete communication graph. Topological attention separately summarizes incoming and outgoing neighbors on the pre-decision graph to condition the weights of individual utilities. It captures the training context of who supplies an agent's inputs and who receives its output (Sections 3.2–3.4). The mixer is training only. At inference, the scheduler continues selecting modes and partners with fixed parameters.

Main-result context. Across nine benchmarks, ACTS gains 3.17/6.19 pp in macro accuracy and uses 19.43%/18.75% fewer inference tokens than Agent Q-Mix on GPT-OSS-120B / DeepSeek-V4-Flash (Tables 1 and 3).

Selected competition-mathematics results (Table 1, ACC %):

BackboneBenchmarkACTSAgent Q-Mix
GPT-OSS-120BBeyond-AIME42.0036.00
DeepSeek-V4-FlashAIME 202570.0060.00
DeepSeek-V4-FlashAIME 202686.6773.33

ACTS ranks first among all evaluated methods in each row.

Figure 4 (p.6) links these mechanisms: training-only value mixing on the left, learned mode and partner selection in the middle, and LLM answer generation and aggregation on the right. Control states and graphs evolve across rounds while inference parameters remain fixed. Section 2.2 situates the mixer within the monotonic and attention-based value-factorization literature, including QMIX and Qatten.

Controlled partner and neighborhood ablations test these mechanisms directly. Larger-team, reduced-data, and frozen-transfer evaluations assess their effectiveness under changes in the execution and training settings. Thank you for encouraging this more precise contribution analysis.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer G6ta — Part 2/4: Controlled evidence and larger teams (Q1)

Thank you for requesting evidence beyond the overall baseline comparison. The controlled ablation isolates directed-neighborhood conditioning within the same mixer architecture, while the larger-team evaluation tests whether the scheduling advantage persists with more candidate partners.

Does choosing the partner matter? Fixing partners lowers GPT-OSS-120B macro accuracy from 75.01% to 72.50% while increasing tokens per question from 5.97K to 8.10K. On DeepSeek-V4-Flash, accuracy falls from 69.43% to 66.10% and tokens rise from 4.03K to 6.20K (Table 2(b)). The extra tokens spent on fixed exchanges do not recover the accuracy of learned routing, supporting the value of choosing partners for the current control state.

Controlled mixer comparison. Variant (a) keeps the other ACTS components and replaces only its mixer with QMIX. To isolate explicit neighborhood conditioning, the new variant (d) zeros only incoming/outgoing neighborhood means. It retains the attention-mixer architecture, agent-index embeddings, and execution-network architecture, and retrains the scheduler (Section 4.3.3, Table 2, Appendix E.7, Table 9).

In both tables, ACC is macro accuracy (%) across nine benchmarks. Tokens are thousands per question.

VariantGPT-OSS ACCGPT-OSS tokensDeepSeek ACCDeepSeek tokens
Full ACTS75.015.9769.434.03
(a) ACTS mixer → QMIX71.837.4163.244.96
(d) Zero mixer neighborhood inputs72.677.4067.116.72

Full ACTS outperforms both variants in accuracy and token use. Removing neighborhood inputs costs 2.34/2.32 pp in accuracy and adds 1.43/2.69K tokens on GPT-OSS/DeepSeek. Since execution-time graph processing and relational routing are retained, this comparison identifies a benefit of explicit directed-neighborhood conditioning during training.

Six-agent evaluation. With N = 6, each agent has five potential partners instead of two. Both methods use 60 training episodes and at most five communication rounds (Section 4.3.4, Appendix E.8).

Backbone, N = 6ACTS ACCAgent Q-Mix ACCACTS tokensAgent Q-Mix tokens
GPT-OSS-120B75.4171.009.9012.72
DeepSeek-V4-Flash70.1964.287.099.32

Across the two backbones, the larger-team results show:

  • Higher accuracy: ACTS gains 4.40 pp on GPT-OSS and 5.91 pp on DeepSeek, computed before rounding.
  • Lower token use: reductions are 22.17% and 23.93%, respectively.
  • Selective participation: Table 10 reports 4.11/4.00 active agents per executed round, versus 5.19/5.10 for Agent Q-Mix (GPT-OSS/DeepSeek).
  • Fewer communication links: directed edges average 5.80/5.58, versus 9.69/9.40, in the same order.

Higher accuracy accompanies more selective participation and communication as the candidate-partner set grows.

The ablation tests the training module on the left of Figure 4 (p.6). The six-agent study evaluates the complete execution path, where the scheduler organizes a larger pool of LLM agents. These tests provide complementary evidence for topology-conditioned learning and the accuracy–token advantage of the resulting policy.

Reporting: Three-agent full-ACTS accuracy is the mean of three training runs. The added ablation and N = 6 accuracy results are computed from benchmark-level correct/total counts.

Revision locations: Sections 4.3.3–4.3.4, Tables 2–3, and Appendices E.7–E.8. The corresponding additions and clarifications are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you for encouraging these more targeted evaluations.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer G6ta — Part 3/4: Matched training and data sensitivity (W1, Q2)

We appreciate your questions about limited training data and fairness. They motivate two complementary checks: matching the training conditions and reducing the number of available training examples while keeping the training budget fixed.

Matched protocol. For each seed, ACTS and Agent Q-Mix use the same training questions in the same order and the same held-out evaluation questions. Both use mixed-task training for 60 episodes, with at most five communication rounds per episode (Section 4.1, p.7, Appendix E.1, p.22). Data preparation checks overlap using dataset sample identifiers, with question text as a fallback key, to maintain training–evaluation separation. Scheduler parameters remain fixed during evaluation. These conditions match the task instances, sample order, training budget, and communication budget.

Reduced-data evaluation. The evaluation uses task-stratified 25% and 50% training-data subsets, with the evaluation set and episode budget held fixed (Section 4.3.4, Table 4, p.10, Appendix E.9.2, Table 13, p.30). The results on GPT-OSS-120B with N = 3 are:

Training-data fractionACTS macro ACC (%)Agent Q-Mix macro ACC (%)ACTS gain (pp)
25%67.9863.32+4.65
50%71.6667.07+4.59
100% reference75.0171.83+3.17

ACTS retains an accuracy advantage at every tested data fraction. With only 25% or 50% of the available training examples, it leads by 4.65 and 4.59 percentage points, respectively, under the same 60-episode budget. Both methods benefit from more training data. ACTS's consistent lead shows that its learned coordination remains effective when fewer examples are available within the evaluated task suite.

Reporting: ACC is the equally weighted average across nine benchmarks. The 25%/50% entries are computed from per-benchmark correct/total counts. The 100% reference is the mean of three independent scheduler-training runs. Gains are computed before rounding.

Revision locations: Section 4.1, Table 4, and Appendices E.1 and E.9.2. The corresponding protocol clarifications and reduced-data results are highlighted in blue\textcolor{blue}{\text{in blue}}.

We appreciate your emphasis on assessing the learned policy under matched data and budgets.

Best regards,
The authors

AuthorsOpenReview ↗

Response to Reviewer G6ta — Part 4/4: Frozen-scheduler transfer across backbones (W1, Q2)

Thank you for suggesting a direct transfer test. To examine whether a learned scheduling policy can be reused, we deploy a GPT-OSS-trained scheduler on DeepSeek with all scheduler parameters frozen (Section 4.3.4, Table 3, p.10, Appendix E.9.1, Table 12, pp.29–30). ACTS and Agent Q-Mix both use this protocol with three agents and the same nine-benchmark suite.

Frozen GPT-OSS → DeepSeekMacro ACC (%)Tokens per question (×10³)ACC drop from native DeepSeek (pp)
ACTS66.675.432.76
Agent Q-Mix61.596.321.65

Frozen ACTS reaches 66.67% accuracy with 5.43K tokens per question, versus 61.59% and 6.32K for Agent Q-Mix. It retains a 5.08-percentage-point accuracy advantage with 14.08% fewer tokens after the LLM backbone changes. ACTS is 2.76 percentage points below its native DeepSeek reference. With the nine benchmarks held fixed, this experiment tests reuse of the learned participation and partner-selection policy across the GPT-OSS-to-DeepSeek backbone change.

Reporting: Transfer ACC is the equally weighted average across nine benchmarks, using each benchmark's correct/total counts. The native-backbone references used to calculate the accuracy drops are means of three independent scheduler-training runs, with a sample SD of 0.45 percentage points in each setting (Appendix E.1.2, Table 6, p.23).

Implications for policy reuse. The scheduler learns how to allocate participation and directed communication, while the frozen LLMs perform the reasoning. The transfer experiment changes the LLM executor while reusing the learned allocation policy. Its accuracy–token advantage in this test complements the reduced-data and larger-team results, demonstrating effective coordination across the evaluated changes in data availability, team size, and backbone.

Revision locations: Section 4.3.4, Table 3, and Appendices E.1.2 and E.9.1. The corresponding additions and clarifications are highlighted in blue\textcolor{blue}{\text{in blue}}.

Thank you again for your time and thoughtful questions. They have helped us clarify how ACTS learns communication schedules and strengthen the evidence for their effectiveness and reuse.

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.