PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems

Xu Shen, Yixin Liu, Yiwei Dai, Yili Wang, Rui Miao, Yue Tan, Shirui Pan, Xin Wang

arXiv:2505.23352v1cs.MAcs.AI

TL;DR

Automated topology design often favors sparse communication but leaves open when sparse or dense topologies support collaboration. The paper analyzes error and insight propagation, then introduces EIB-LEARNER to fuse sparse and dense views; experiments report improved performance, robustness, and communication efficiency across reasoning, math, and code tasks.

  • Problem

    Automated topology-design methods often favor sparsification, while why and when sparse or dense topologies help collaboration remains an open question.

    Source: §1 · p. 2

  • Method

    EIB-LEARNER uses causal analysis of information propagation and query-aware fusion of sparse and dense graph views to design communication topologies.

    Source: §1 · p. 2

  • Results

    Across reasoning, math, and code tasks, EIB-LEARNER consistently improves performance and robustness while reducing communication overhead.

    Source: §6 · p. 9

  • Takeaways & Limitations

    Moderate sparsity is identified as a balance that suppresses error propagation while preserving beneficial insight diffusion.

    Source: §1 · p. 2

    Evaluation is limited to reasoning, math, and code tasks, and the design relies on predefined agent roles and manually crafted prompts.

    Source: Limitation · p. 9

Abstract

from arXiv · show

The communication topology in large language model-based multi-agent systems fundamentally governs inter-agent collaboration patterns, critically shaping both the efficiency and effectiveness of collective decision-making. While recent studies for communication topology automated design tend to construct sparse structures for efficiency, they often overlook why and when sparse and dense topologies help or hinder collaboration. In this paper, we present a causal framework to analyze how agent outputs, whether correct or erroneous, propagate under topologies with varying sparsity. Our empirical studies reveal that moderately sparse topologies, which effectively suppress error propagation while preserving beneficial information diffusion, typically achieve optimal task performance. Guided by this insight, we propose a novel topology design approach, EIB-leanrner, that balances error suppression and beneficial information propagation by fusing connectivity patterns from both dense and sparse graphs. Extensive experiments show the superior effectiveness, communication cost, and robustness of EIB-leanrner.

1 Introduction

Automated topology design often prioritizes sparse graphs without explaining when sparse or dense communication helps collaboration. The paper analyzes information propagation and proposes a topology learner that balances error suppression with beneficial insight diffusion.

1 Introduction

  • The paper identifies a gap in automated topology design: sparsification is often treated as an end rather than a question of when sparse or dense graphs help collaboration.Source: §1 · p. 2
  • Moderately sparse topologies can suppress error propagation while preserving beneficial insight diffusion, making them the identified ideal balance for collective decisions.Sparse topologies may suppress useful insights, whereas dense topologies enable their dissemination but also aggressively propagate individual errors.Source: §1 · p. 2
  • EIB-LEARNER uses dual-view graph neural networks and query-aware fusion to combine connectivity patterns learned from sparse and dense graphs.Its two views simulate sparse-graph error suppression and dense-graph insight propagation.Source: §1 · p. 2
  • The paper uses causal counterfactual analyses to assess how agents affect collective decisions across communication topologies with varying sparsity.Source: §1 · p. 2
  • Across six MAS decision-making benchmarks, EIB-LEARNER achieves superior accuracy with reduced communication cost and better robustness than existing baselines.Source: §1 · p. 2

2 Problem Definition

The problem is to choose a communication topology that maximizes collective-output utility for a task query. The system represents communication as a directed acyclic graph and runs agents in topological order over multiple interaction rounds.

2 Problem Definition

  • A communication topology is a directed acyclic graph whose nodes are LLM agents and whose directed edges specify communication links.Connectivity is represented by an adjacency matrix.Source: §2 · p. 2
  • Agents execute in topological order over K interaction rounds, using their query and neighbor outputs to generate responses before a final aggregation.Aggregation can use majority voting, a designated agent, or the last agent in execution order; K may be predefined or adaptively stopped.Source: §2 · pp. 2–3
  • The design objective is to select a feasible topology that maximizes the utility of the collective output for a given task query.The utility evaluates correctness or overall output quality; feasible topologies are DAGs satisfying acyclicity and predefined execution constraints.Source: §2 · p. 3

3 Analysis of Communication Topology

The section analyzes how communication topology affects agent-level information propagation and collective decision-making. It uses counterfactual measures and empirical comparisons across sparsity levels and topology types.

3.1 Causal Metric of Agent Impact

  • CAPE measures whether changing one agent’s output changes the final answer, while TCTE extends intervention sensitivity to the topology level.TCTE normalizes agent influence using degree.Source: §3.1 · p. 3

3.2 Empirical Analysis

  • The empirical study uses six GPT-3.5 agents on MMLU, with separate sets of 500 originally correct and 500 originally incorrect questions.The two sets support simulations of detrimental and beneficial agent-level interventions.Source: §3.2 · p. 4

3.2.1 Analysis of Error Propagation Effect

  • With errors injected into originally correct cases, denser graphs are more susceptible to error propagation; Chain reduces TCTE by up to 19%.Full reaches peak TCTE, whereas sparser structures mitigate localized errors.Source: §3.2.1 · p. 4

3.2.2 Analysis of Insight Propagation Effect

  • With correct insights injected into originally incorrect cases, denser graphs strengthen insight propagation; Full increases TCTE by 10.5%.Chain has the lowest TCTE, indicating poor transmission of correct insights.Source: §3.2.2 · pp. 4–5

3.2.3 Effectiveness Analysis

  • Moderate sparsity gives the best task performance by suppressing erroneous information while preserving beneficial insight propagation.The proposed agent-level pattern is dense connectivity around accurate agents and sparse links for error-prone agents.Source: §3.2.3 · p. 5

3.2.4 Analysis of Existing Topologies

  • AgentDrop and AgentPrune do not outperform random graphs with moderate sparsity, despite sparse graphs resisting errors and dense graphs amplifying helpful insight.The comparison motivates topology design that explicitly balances both propagation effects.Source: §3.2.4 · p. 5

4 Methodology

EIB-LEARNER designs query-specific MAS communication topologies by modeling error suppression and insight propagation in complementary sparse and dense views, then fusing them. Its GNN-based simulators estimate agent connectivity, while a query-aware gate combines the estimates into a topology.

4 Methodology

  • EIB-LEARNER co-trains sparse- and dense-view GNN simulators, then adaptively blends their learned inter-agent coefficients to construct robust topologies.The sparse view models error suppression and the dense view models insight propagation.Source: §4 · p. 5

4.1 GNN-based Propagation Simulators

  • Agent-role information and the query are incorporated into GNN node features to model each agent’s likely beneficial or error-prone contribution for that query.Message aggregation models agent communication, while feature transformation models agents responding to a query.Source: §4.1 · pp. 5–6
  • The two simulators use sparse and dense graph views to model error suppression and insight propagation, respectively.The framework generates complementary topological representations for the two propagation effects.Source: §4.1 · p. 6, Figure 3 · p. 6

4.2 Error-Insight Balanced Topology Learner

  • An inner-product decoder estimates pairwise connectivity from propagated information, producing coefficients for the sparse and dense views.A query-aware MLP outputs softmax fusion weights that determine each view’s contribution to the final coefficient matrix; the matrix is used to sample a discrete topology.Source: §4.2 · pp. 6–7
  • The model optimizes topology utility based on final-answer correctness using policy gradients because downstream LLM reasoning makes the utility function nondifferentiable.The final answer’s task-specific utility serves as the reward.Source: §4.2 · p. 7

5 Experiments

Across six benchmarks, EIB-LEARNER reports the highest average accuracy among the evaluated methods, with lower token cost than predefined dense topologies and robustness to prompt attacks. Ablations indicate that its dense branch, sparse branch, and fusion module each contribute to performance.

5.1 Experimental Setup

  • The evaluation covers six benchmarks spanning general reasoning, mathematical reasoning, and code generation, with single-agent, fixed-topology, LLM-Debate, and automated-design baselines.Experiments use GPT-4o, three interaction rounds, and task-specific agent counts.Source: §5.1 · p. 7

5.2 Performance Comparison

  • 91.38% average accuracy across six benchmarks; EIB-LEARNER reports a 1.34% average improvement over recent best topology-optimization methods.The reported average is higher than the 90.04% achieved by G-designer.Source: §5.2 · p. 7, Table 1 · p. 8

5.3 Ablation Study

  • Removing the dense branch causes the largest ablation drop, including a 4.1% accuracy decrease on GSM8K; removing the sparse branch or fusion also degrades performance.The study evaluates these components on MMLU, GSM8K, and HumanEval.Source: §5.3 · p. 8, Table 2 · p. 8

5.4 Token Efficiency

  • On MMLU and GSM8K, EIB-LEARNER has token cost comparable to G-designer while delivering significantly higher accuracy, and it reduces token consumption versus fully connected graphs and LLM-Debate.The paper attributes this to eliminating redundant interactions while preserving critical information flow.Source: §5.4 · p. 8, Figure 4 · p. 8

5.5 Robustness Analysis

  • Under single-agent prompt injection, EIB-LEARNER’s accuracy drops by 1.24%, while chain and tree topologies drop by up to 11.8%.The paper attributes its robustness to mitigating error propagation through bottlenecks while promoting consensus among densely connected reliable agents.Source: §5.5 · p. 8, Figure 4 · p. 8

6 Conclusion

The paper introduces EIB-LEARNER, a topology optimization framework grounded in causal analysis of information propagation. It reports improved performance, robustness, and communication efficiency across reasoning, math, and code tasks.

6 Conclusion

  • EIB-LEARNER adaptively fuses dense and sparse graph views to form task-specific topologies, balancing error suppression with insight preservation.The framework is grounded in causal analysis of information propagation.Source: §6 · p. 8
  • Experiments across reasoning, math, and code tasks show that EIB-LEARNER improves performance and robustness while reducing communication overhead.Source: §6 · p. 9

Limitation

The evaluation covers reasoning, math, and code generation, while the framework uses predefined agent roles and manually crafted prompts. The authors identify broader task domains and automated role and prompt design as future work.

Limitation

  • The evaluation is limited to reasoning, math, and code generation tasks.Source: Limitation · p. 9
  • The design relies on predefined agent roles and manually crafted prompts, which may limit adaptability in unfamiliar or evolving scenarios.The authors suggest automatically discovering roles and optimizing prompts as future work.Source: Limitation · p. 9

Ethics Statement

The paper reports no involvement of human subjects, animals, or environmentally sensitive materials. Its background describes LLM multi-agent systems and graph-based communication topologies, including learnable sparse structures.

Ethics Statement

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

A.1 LLM-based Multi-Agent Systems

  • LLM-based multi-agent systems use strategies such as sequential pipelines, debate, and centralized planning to organize inter-agent interaction.Source: §A.1 · p. 12

A.2 MAS as Graphs

  • Learnable communication graphs can improve efficiency and scalability, but prior methods generally treat sparsity as cost-saving without analyzing when sparse or dense topologies improve outcomes.Source: §A.2 · p. 12

B Pseudo-code for EIB-LEARNER

EIB-LEARNER uses a lightweight encoder, two GNNs, and gated fusion to design communication topologies. In systems with 5 to 9 agents, it outperforms G-Designer while maintaining comparable computational cost and runtime.

B Pseudo-code for EIB-LEARNER

  • EIB-LEARNER takes the current query as input and comprises a lightweight language encoder, two GNNs, and a gated fusion network.During training, a subset of queries is used to construct topologies, followed by multi-agent decision-making with those topologies.Source: B Pseudo-code for EIB-LEARNER · p. 12

C Scalability in the number of agents

  • 5.11% improvement in accuracy with about half the token cost of a fully connected topology; EIB-LEARNER maintains comparable computational cost and runtime to G-Designer.The comparison covers systems with 5 to 9 agents; Table 3 compares accuracy, time, and cost across agent configurations.Source: C Scalability in the number of agents · p. 12, Table 3 · p. 12
  • Figure 5 presents case studies of EIB-LEARNER-designed communication topologies on HumanEval and GSM8K.Source: Figure 5 · p. 13

D Dataset Statistic

This section presents dataset statistics and follows the same experimental setup as G-Designer (Zhang et al., 2025b).

D Dataset Statistic

E Case Study

The case studies illustrate how EIB-LEARNER assigns communication roles and verification pathways for HumanEval and GSM8K. These pathways are described as supporting correct reasoning while suppressing misleading signals.

E Case Study

  • In HumanEval, a project manager passes planning through design and implementation, followed by verification and bug-fixing feedback.The Test Analyst and Bug Fixer form a feedback loop described as suppressing local errors while ensuring functional accuracy.Source: E Case Study · p. 13
  • For GSM8K, the Mathematical Analyst communicates with the Math Solver, while the Programming Expert provides structural parsing to the Inspector.The Inspector checks the calculation logic to avoid propagation of misunderstandings.Source: E Case Study · p. 13
  • The case-study topologies are described as enabling correct reasoning to propagate efficiently while suppressing misleading signals through targeted verification.Source: E Case Study · p. 13
  • Algorithm 1 is titled as the EIB-LEARNER workflow and lists a query, NodeEncoder, two GNNs, a gating function, and learning rate α as inputs.Source: E Case Study · p. 14
Loading 2505.23352v1…