PaperlayerView exact revision on arXiv

Source-linked AI summary

GoAgent: Group-of-Agents Communication Topology Generation for LLM-based Multi-Agent Systems

Hongjiang Chen, Xin Zheng, Yixin Liu, Pengfei Jiao, Shiyuan Li, Huan Liu, Zhidong Zhao, Ziqi Xu, Ibrahim Khalil, Shirui Pan

arXiv:2603.19677v1cs.LGcs.AIcs.MA

TL;DR

Node-centric topology generation leaves collaborative groups implicit, despite their role in task coordination. GoAgent generates topologies from groups and compresses inter-group communication; across six benchmarks it reports state-of-the-art accuracy and reduced communication overhead.

  • Problem

    Existing node-centric topology methods leave collaborative group structures implicit, while topology design must support effective coordination and efficient token usage.

    Source: §1 · pp. 1–2

  • Method

    GoAgent treats task-relevant collaborative groups as atomic units, autoregressively selects and connects them, and uses a conditional information bottleneck to compress communication.

    Source: §1 · p. 2, §2.1 · p. 3

  • Results

    Across six benchmarks, GoAgent consistently achieves state-of-the-art accuracy, improves structural robustness, and reduces communication overhead.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    GoAgent demonstrates a group-centric approach that adaptively forms task-specific topologies while suppressing task-irrelevant noise.

    Source: §6 · p. 8

    GoAgent relies on an offline LLM-generated group pool and was evaluated primarily on static reasoning tasks, leaving its effectiveness in dynamic interactive environments unestablished.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language model (LLM)-based multi-agent systems (MAS) have demonstrated exceptional capabilities in solving complex tasks, yet their effectiveness depends heavily on the underlying communication topology that coordinates agent interactions. Within these systems, successful problem-solving often necessitates task-specific group structures to divide and conquer subtasks. However, most existing approaches generate communication topologies in a node-centric manner, leaving group structures to emerge implicitly from local connectivity decisions rather than modeling them explicitly, often leading to suboptimal coordination and unnecessary communication overhead. To address this limitation, we propose GoAgent (Group-of-Agents), a communication topology generation method that explicitly treats collaborative groups as the atomic units of MAS construction. Specifically, GoAgent first enumerates task-relevant candidate groups through an LLM and then autoregressively selects and connects these groups as atomic units to construct the final communication graph, jointly capturing intra-group cohesion and inter-group coordination. To mitigate communication redundancy and noise propagation inherent in expanding topologies, we further introduce a conditional information bottleneck (CIB) objective that compresses inter-group communication, preserving task-relevant signals while filtering out redundant historical noise. Extensive experiments on six benchmarks demonstrate the state-of-the-art performance of GoAgent with 93.84% average accuracy while reducing token consumption by about 17%.

1 Introduction

Communication topology shapes how multi-agent systems coordinate, but node-centric construction leaves collaborative groups implicit. GoAgent instead builds around task-relevant groups and uses conditional compression to reduce redundant communication.

1 Introduction

  • Communication topology is central to agent interaction, information sharing, and coordination, and prior studies report that organizational structure affects system performance more than individual agent capabilities.Earlier static chains, trees, and fully connected graphs were inflexible across task requirements; later template-based methods remained bounded by their initial agents and connections.Source: §1 · p. 1
  • Node-centric generation makes local agent-and-edge decisions without a macroscopic workflow view, leaving tightly coupled task groups to emerge implicitly and risking disjointed coordination.Unconstrained connectivity can also create redundant message passing and task-irrelevant historical noise.Source: §1 · pp. 1–2
  • GoAgent treats collaborative groups as atomic construction units, then autoregressively selects and connects task-relevant groups to preserve intra-group cohesion and coordinate between groups.Its conditional information bottleneck compresses inter-group communication, retaining task-relevant signals while filtering redundant historical noise.Source: §1 · p. 2
  • Experiments on six benchmarks show state-of-the-art accuracy alongside reduced token consumption.The introduction identifies both coordination quality and token efficiency as design goals for topology generation.Source: §1 · pp. 1–2

2 Preliminaries

The preliminaries formulate topology generation as constructing a group-level communication graph for a task query. They define collaborative groups and describe conditional information bottleneck as compression conditioned on a target-relevant variable.

2.1 Problem Formulation

  • A collaborative group is a subset of agents with predefined internal communication relations, and the final agent topology combines those relations with dependencies between groups.A group-level edge indicates that information from one group can be transmitted to another.Source: §2.1 · p. 3
  • Given a task query, the objective is to generate a communication structure that coordinates agents to solve the task, with groups selected sequentially and connected to previously constructed groups.The generation distribution models group selection and inter-group edge prediction as separate terms.Source: §2.1 · pp. 2–3

2.2 Information Bottleneck

  • The information bottleneck compresses signals while preserving information relevant to a target; conditional information bottleneck performs this compression conditioned on a variable.Conditioning selectively filters noise while preserving patterns dictated by that condition.Source: §2.2 · p. 3

3 Methodology

GoAgent constructs task-specific multi-agent topologies with collaborative groups as atomic units, combining group selection, inter-group edge prediction, and task-conditioned compression. It trains on curated minimal successful graphs and generates graphs autoregressively at inference.

3 Methodology

  • GoAgent uses collaborative groups as generative atomic units, reducing the search space and aligning topology construction with human-like organizational structures.Source: §3 · p. 3

3.1 Task Encoding and Group Discovery

  • A pre-trained sentence encoder and feed-forward network map each task query to a global representation that guides subsequent generation.Source: §3.1 · pp. 3–4
  • An LLM proposes task-relevant groups with named expertise, roles, and fixed intra-group topologies, which the generator uses to model inter-group dependencies.The candidate groups are encoded into an embedding matrix and augmented with END tokens.Source: §3.1 · p. 4

3.2 Autoregressive Group-Centric Generation

  • A GRU aggregates previously generated group embeddings, while a dynamic gate fuses historical context with the task representation to guide group and edge generation.Source: §3.2 · p. 4

3.3 Conditional Information Bottleneck

  • The CIB layer compresses historical group and edge features conditioned on the task representation, retaining predictive information while limiting task-irrelevant signals.The method assumes a Markov chain from task representation through historical features and compressed features to the prediction, and uses a task-conditioned prior.Source: §3.3 · p. 5

3.4 Training and Inference Strategy

  • Training uses teacher forcing on curated trajectories from successful graphs filtered for minimal groups and edges; inference selects groups autoregressively until END and samples incoming edges.The joint objective combines group and edge prediction losses with KL regularization at both levels; KL weights are warmed up during early training.Source: §3.4 · pp. 5–6

4 Experiments

Across six benchmarks, GoAgent reports state-of-the-art accuracy, lower token consumption, and stronger robustness; ablations and case studies examine its components and generated collaboration graphs.

4 Experiments

  • Experiments cover six reasoning and coding benchmarks, compare single-agent, fixed-topology, and learnable-topology baselines, and train using only 40 or 60 queries per dataset.Implementation uses GPT models, primarily gpt-4o-2024-08-06, and sets the maximum interaction round to three.Source: §4 · p. 6

4.1 Performance Comparison

  • GoAgent achieves state-of-the-art accuracy across six benchmarks, with larger gains on complex tasks including MMLU (1.96% ↑) and HumanEval (2.47% ↑) against ARG-Designer and EIB-LEARNER.The comparison is reported in Table 1.Source: §4.1 · p. 6, Table 1 · p. 7

4.2 Ablation Study

  • Without Group, MMLU accuracy decreases 2.61%; without CIB, MMLU accuracy drops 3.27%.The authors describe group-level generation and structural noise filtering as complementary.Source: §4.2 · p. 7, Table 2 · p. 7

4.3 Token Efficiency

  • GoAgent reduces token consumption by 17% versus the SOTA baseline while maintaining high accuracy.The paper attributes the reduction to group-level construction and CIB filtering irrelevant inter-group edges.Source: Figure 3 · p. 7, §4.3 · p. 7

4.4 Robustness Analysis

  • Under a simulated single-agent system-prompt attack, GoAgent maintains 89.54% accuracy, while node-centric G-Designer and ARG-Designer suffer sharp performance drops.The CIB filters incoming features conditioned on the task representation, restricting localized hallucination spread across group boundaries.Source: Figure 3 · p. 7, §4.4 · p. 8

4.5 Case Study

  • In representative MMLU cases, GoAgent builds graphs with fewer agents and messages than ARG-Designer, lowering token usage without sacrificing accuracy.ARG-Designer’s graph density increases sharply with task complexity, whereas GoAgent selects from an explicitly constructed group pool.Source: Figure 4 · p. 8, §4.5 · p. 8

5 Related Work

Prior topology methods move from fixed structures to learned pruning and autoregressive generation, but remain node-centric; GoAgent instead builds around collaborative groups.

LLM-based Multi-Agent Systems

  • Existing adaptive approaches include template-based pruning and ARG-Designer’s autoregressive construction, but all treat individual agents as atomic units and can miss cohesive structures or generate redundant edges.The related-work discussion contrasts these methods with earlier static chains and other fixed architectures.Source: §5 · p. 8

Group-Aware Graph Generation

  • GoAgent applies higher-order graph-generation ideas by using LLM-derived collaborative groups as atomic building blocks, preserving intra-group cohesion and simplifying inter-group coordination.The cited graph literature includes hierarchical networks, diffusion-based methods, and scalable higher-order frameworks.Source: §5 · p. 8

6 Conclusion

GoAgent uses group-centric topology generation and a conditional information bottleneck to form task-specific topologies while suppressing irrelevant noise. Across six benchmarks, it reports state-of-the-art accuracy, improved structural robustness, and reduced communication overhead.

6 Conclusion

  • GoAgent treats collaborative groups as generative units and uses a CIB to compress historical signals, with experiments reporting state-of-the-art accuracy, improved structural robustness, and reduced communication overhead across six benchmarks.The method adaptively forms task-specific topologies while suppressing task-irrelevant noise.Source: §6 · p. 8

Limitations

The authors identify limits in group-pool flexibility and in the settings evaluated so far.

Limitations

  • GoAgent relies on an offline LLM-generated pool of collaborative groups and cannot synthesize novel groups or roles during inference when a task requires expertise outside that pool.The predefined pool reduces the search space and ensures intra-group cohesion, but constrains flexibility.Source: Limitations · p. 9
  • Evaluation primarily covers static reasoning tasks, leaving effectiveness in dynamic interactive settings such as embodied AI and multi-agent reinforcement learning to future investigation.Source: Limitations · p. 9

Ethics Statement

The paper reports no ethical risks or conflicts of interest. GoAgent extends information bottleneck to topology generation, with a lightweight generator whose computational overhead is negligible relative to the underlying LLM inference.

Ethics Statement

  • The authors report no ethical risks or conflicts of interest and state that the research involves no human subjects, animals, or environmentally sensitive materials.Source: Ethics Statement · p. 9

A More Related Work

  • GoAgent extends information bottleneck to MAS topology generation by conditioning compression on task representations to filter redundant historical communication noise.This builds on prior uses of information bottleneck for graph learning and LLM-based MAS.Source: A More Related Work · p. 12

B Algorithm and Complexity Analysis

  • The generator’s time complexity is O(T^2d^2 + TKd), and its overhead is negligible because T and K are small relative to LLM inference cost.The paper gives typical T < 3 and K = 16.Source: B Algorithm and Complexity Analysis · p. 12
  • Generator space complexity is O(d^2 + Kd), from lightweight GRUs, MLPs, CIB projections, and candidate-group embeddings.The paper characterizes the generator as parameter-efficient and easy to integrate alongside LLMs.Source: B Algorithm and Complexity Analysis · p. 12

C Dataset Details

The paper describes dataset statistics and evaluates on six diverse benchmarks, then compares GoAgent with single-agent, fixed-topology, and learning-based topology methods.

C Dataset Details

D Baseline Details

  • The single-agent baselines are Vanilla prompting, Chain-of-Thought, and Self-Consistency using five reasoning paths and majority voting.Self-Consistency samples five CoT paths before selecting the majority answer.Source: D Baseline Details · p. 12
  • Fixed-topology comparisons include Chain, Tree, Complete, Random, and LLM-Debate structures.These span sequential, hierarchical, fully connected, randomly generated, and debate-based designs.Source: D Baseline Details · p. 12
  • Learning-based comparisons include AgentPrune, Agent-Dropout, G-Designer, and EIB-LEARNER; the methods are trained separately per dataset.The passage also begins describing ARG-Designer as an autoregressive topology method that adds individual agents and pairwise connections without predefined templates.Source: D Baseline Details · pp. 12–14

E Parameter Sensitivity Analysis

The MMLU sensitivity analysis shows that CIB strength affects performance, with too little or too much compression associated with different topology problems. The reported optimum is asymmetric between group and edge prediction.

E Parameter Sensitivity Analysis

  • Figure 5 presents MMLU performance as a heatmap over varying βg and βe values.Source: Figure 5 · p. 14, E Parameter Sensitivity Analysis · p. 14
  • βe = 0.0 leaves the bottleneck ineffective, allowing historical noise through and producing over-connected graphs with high token costs and lower accuracy.Source: E Parameter Sensitivity Analysis · p. 14
  • βe > 0.5 over-compresses features, discarding necessary historical context and yielding under-connected, poorly coordinated topologies.Source: E Parameter Sensitivity Analysis · p. 15
  • βg ≈ 0.0 and βe ≈ 0.3 is the reported optimal configuration, indicating greater benefit from compression for edge prediction than group prediction.The authors relate group prediction to retaining historical context to avoid duplicate roles, while edge prediction is susceptible to redundant connectivity patterns.Source: E Parameter Sensitivity Analysis · p. 15

F Implementation Details

GoAgent is trained with a small number of queries per dataset and a warm-up schedule for CIB strength. Its prompt templates generate diverse query-topology pairs, while its group pool encodes multiple role configurations and collaboration patterns.

F Implementation Details

  • Training uses AdamW for 100 epochs with batch size 40, and linearly warms up CIB strength over the first 10 epochs.Only 40 or 60 training queries per dataset are used because the generator learns structural collaboration patterns rather than task-specific reasoning.Source: F Implementation Details · p. 15

F.1 LLM Prompt Templates

  • GPT-4o at temperature 0.7 generates query-topology pairs across domains including coding, mathematics, multiple-choice questions, knowledge Q&A, medical consultation, and psychology.The prompts propose MAS configurations using roles from the group pool and Chain, Star, or FullConnected topologies.Source: §F.1 · p. 15

F.2 Group Role Prompts

  • The group pool comprises 16 configurations derived from four base roles and five topology patterns, covering standalone, chain, star, and fully connected collaboration structures.Internal group topologies are directed acyclic graphs, with information flowing from upstream to downstream roles.Source: §F.2 · p. 15
  • Example groups assign Analyst and Solver outputs to an Inspector for validation, or pass analysis through a Solver before inspection.The full-validation pattern also sends Analyst output to both Solver and Inspector before the Inspector checks the result against both sources.Source: §F.2 · p. 16

Related papers

Loading 2603.19677v1…