PaperlayerView exact revision on arXiv

Source-linked AI summary

BusMA: A Bus Communication Substrate for Multi-Agent Systems

Yanwen Peng, Delvin Ce Zhang, Xi Wang, Nikolaos Aletras

arXiv:2609.15054v2cs.AI

TL;DR

HMW and RMP limit direct communication among Workers and route information through a central hub. BusMA provides a shared Bus for peer messaging and Chair-led coordination; experiments report that it outperforms HMW and RMP across tasks.

  • Problem

    HMW and RMP restrict Workers from directly consulting peers and route communication through a manager or router.

    Source: §1 · p. 1

  • Method

    BusMA gives Workers direct peer messaging through a shared Bus, while an Act-disabled Chair monitors shared history and coordinates interactions.

    Source: §1 · p. 2, §3.2 · p. 4, §3.3 · p. 5

  • Results

    Across 13 tasks, BusMA consistently outperforms state-of-the-art HMW and RMP methods.

    Source: §1 · p. 2, §7 · p. 8

  • Takeaways & Limitations

    BusMA improves collaboration through interaction but does not enhance an LLM's underlying capabilities.

    Source: Limitations · p. 9

    Shared-memory errors can bias the Chair's synthesis, particularly when tools provide incomplete or wrong evidence and Workers cannot verify unfamiliar tools.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Multi-Agent (MA) systems are effective at solving complex tasks that demand planning, tool use, and the synthesis of evidence from multiple sources. Existing systems typically adopt Hierarchical Manager-Worker (HMW) or Router-based Message Passing (RMP) structures as their communication protocol. However, these designs restrict agent autonomy: Worker agents cannot directly consult specific "peers", and misrouted messages can propagate errors. Inspired by bus architectures in computer systems, we propose BusMA, a communication framework that allows any agent to address other agents through a shared channel, i.e., the Bus. It consists of agent registration, message routing, and shared memory management components. Worker agents, each equipped with tools, have their own local memory and can reason, act (tool usage), and communicate by posting shared messages with specific intents. We introduce four intents: discussion, challenge, guidance, and request for explanation, which support fine-grained communication among agents. A Chair agent monitors the shared memory to coordinate interactions and facilitate convergence among Workers. To evaluate the effectiveness of BusMA, we conduct extensive experiments with two frontier LLMs across 13 tasks spanning visual reasoning, mathematical reasoning, and knowledge retrieval. The results demonstrate that BusMA consistently outperforms state-of-the-art HMW and RMP methods.

1 Introduction

BusMA replaces hub-mediated communication with a shared channel for direct peer messaging. Its experiments report gains over HMW and RMP, with larger collaboration gains on peer-verifiable tasks than on perceptual tasks.

1 Introduction

  • HMW and RMP restrict Workers to passive execution without direct peer communication, while their central manager or router can become an error-propagating cognitive bottleneck.The cited limitations include barriers to critique and error correction as well as forcing information through a single hub.Source: §1 · p. 1
  • BusMA lets agents directly address specific peers through a shared Bus rather than waiting for assignments or router mediation.The design draws inspiration from distributed systems and computer bus architecture.Source: §1 · p. 2
  • Workers keep local memory and communicate through Bus messages with four intents, while a Chair monitors shared memory and issues follow-ups to facilitate convergence.The intents are discussion, challenge, guidance, and request for explanation.Source: §1 · p. 2
  • Across 13 datasets, BusMA outperforms competitive HMW and RMP frameworks, with larger collaboration gains on peer-verifiable tasks and competitive performance on perceptual tasks.The analysis also finds that fine-grained role division is not uniformly beneficial and Worker heterogeneity can let errors propagate undetected.Source: §1 · p. 2

2 Related Work

Prior work uses LLM agents for diverse tasks, but single-agent systems remain brittle on long-horizon, interdependent work. Multi-agent systems commonly coordinate through centralized HMW or RMP hubs, motivating more flexible communication strategies.

2.1 LLM-based Agents.

  • LLM agents augment language models with planning, tools, knowledge bases, memory, or reflection, and are applied across programming, system interaction, and scientific discovery.Prior evaluations report systematic single-agent failures in planning, decision-making, and instruction following on long-horizon, interdependent tasks.Source: §2.1 · p. 2

2.2 LLM-based MA Systems.

  • HMW managers and RMP routers mediate coordination and information flow, while restricted communication and possible error propagation motivate more flexible strategies.Examples include AgentVerse, ChatDev, and OctoTools for HMW, and AutoGen for RMP.Source: §2.2 · pp. 2–3

3 The BusMA Framework

BusMA combines tool-using Workers, an Act-disabled Chair, and a shared Bus for on-demand peer collaboration. Workers keep private reasoning local while the Bus carries messages and shared history for coordination and synthesis.

3 The BusMA Framework

  • BusMA comprises tool-using Workers with local memory, an Act-disabled Chair, and a Bus for registration, message routing, and shared-memory management.The Chair begins the task, monitors progress, and can issue follow-up calls or complete the response.Source: §3 · p. 3

3.1 Worker Agent

  • Workers alternate reasoning with tool use or peer messaging, retaining private reasoning and tool outputs in local memory across iterations.A Call sends a shared message to a chosen agent and suspends the calling Worker; the Worker iteration also ends at its maximum limit.Source: §3.1 · pp. 3–4
  • Workers frame messages with discussion, request-for-explanation, challenge, or guidance intents, pairing each with evidence and a focused question.Challenge lets one agent question another's intermediate results for re-examination through deliberation.Source: §3.1 · p. 4

3.2 Chair Agent

  • The Act-disabled Chair coordinates from the task entry point without upfront task decomposition, while Workers can exchange messages directly without routing through it.On reactivation, it checks for unresolved disagreements, missing evidence, or unanswered questions, then follows up or synthesizes a final response.Source: §3.2 · p. 4

3.3 Communication Bus

  • The Bus registers unique agent addresses, routes calls into per-agent FIFO queues, and preserves shared messages in routing order for the Chair's global assessment.Each receiver consumes messages in arrival order, while the shared history gives the Chair access to the communication record when a Worker messages it.Source: §3.3 · pp. 4–5

3.4 Overall Information Flow in BusMA

  • Workers keep intermediate evidence and tentative conclusions private, publicizing only actionable content with an intent and supporting evidence when they call the Bus.The Chair consults shared history to identify conflicts or gaps and issue targeted follow-ups; the authors hypothesize this separation reduces overhead while preserving coordination coherence.Source: §3.4 · p. 5

4 Experimental Setup

The evaluation covers breadth across visual, mathematical, and knowledge-retrieval tasks, plus depth on complex GAIA questions. It compares general-purpose multi-agent frameworks and specialized GAIA systems under controlled implementation settings.

4.1 Benchmarks

  • Twelve diversity-oriented datasets cover visual reasoning, mathematical reasoning, and knowledge retrieval, with 200 instances per dataset; GAIA adds 165 questions requiring file parsing, web browsing, and code execution.The study reports exact-match accuracy on GAIA.Source: §4.1 · p. 5

4.2 Baselines

  • For diversity tasks, BusMA is compared with OctoTools, SmolAgents, LangGraph, and AutoGen; GAIA comparisons include MagenticOne, OpenDeepResearch, and Gemini function-calling baselines.Table 3 reports accuracy across tasks and models, while Table 4 reports GAIA performance.Source: §4.2 · pp. 5–6, Table 3 · p. 6, Table 4 · p. 6

4.3 Implementation Details

  • Experiments use DeepSeek-V3 and Gemini-2.5-Flash, route DeepSeek visual calls to Gemini-2.0-Flash, and activate only one agent at a time to isolate communication structure from parallelism.Task-specific Workers include Web, ImageQA, and Code agents; GAIA uses adapted Browser Agent and file-analysis tools.Source: §4.3 · p. 6

5 Results

BusMA outperforms strong baselines across the diversity benchmarks and on GAIA, with especially strong results on mathematical and knowledge-intensive tasks. Its relative gains vary by task and backbone, and retrieval errors can still mislead coordination.

5.1 Diversity Benchmarks

  • BusMA surpasses the best diversity-task baseline, OctoTools, by 4.8% with DeepSeek-V3 and 2.5% with Gemini-2.5-Flash, ranking first or second on seven of eight visual task–backbone combinations.On visual reasoning, the improvements are larger with DeepSeek-V3 and more modest with Gemini-2.5-Flash.Source: §5.1 · p. 6
  • BusMA achieves the best results on Game of 24, Omni-MATH, and CLEVR-Math across both backbones, and leads on MathVista with Gemini-2.5-Flash at 79.0%.The authors suggest its communication intents support iterative verification and may mitigate error propagation.Source: §5.1 · p. 6
  • BusMA achieves the best results on MMLU-Pro, SciFIBench, and HotpotQA with both backbones, but ranks second on GPQA with DeepSeek-V3.The authors associate the knowledge-task advantage with local Worker memory and Chair synthesis of shared information; limited retrieval support can leave Worker errors uncorrected.Source: §5.1 · pp. 6–7
  • Gemini-2.5-Flash outperforms DeepSeek-V3, while BusMA’s relative improvement is larger with DeepSeek-V3.The authors suggest BusMA may yield greater gains when the base model is weaker.Source: §5.1 · p. 7

5.2 GAIA Benchmark

  • 48.5% overall GAIA accuracy represents a 4.3% improvement over OpenDeepResearch, with BusMA reaching 60.3%, 47.6%, and 26.9% on Levels 1, 2, and 3.The advantage is 4.6% on Level 2 and 7.7% on Level 3; the same Browser Agent is used as OpenDeepResearch.Source: §5.2 · p. 7

6 Analysis

Ablations and communication trajectories examine BusMA’s components and interactions, while a GPQA failure illustrates how specialized-tool retrieval errors can evade cross-checking.

6.1 Ablation Study

  • Replacing the Chair with an HMW Manager and disabling Worker communication reduces accuracy by 1.5 to 7.0 percentage points across tasks.The Manager setting prevents Workers from initiating peer-to-peer critique or error correction.Source: §6.1 · p. 7
  • With SmolAgents Workers substituted but the Chair and Bus unchanged, performance stays within 1.5% of BusMA and exceeds SmolAgents by 4.0% to 10.0% across tasks.The authors infer that communication accounts for most of BusMA’s improvement and that the design is robust to Worker implementation.Source: §6.1 · p. 7

6.2 Qualitative Analysis

  • On the Omni-MATH trajectory, the Chair requests an explanation and guides validation, while the Code Agent challenges hypotheses; their interaction reaches the correct solution after initial errors.The corresponding BusMA trajectory is shown in Figure 4; Table 8 contains the full trajectory.Source: §6.2 · pp. 7–8, Figure 4 · p. 8
  • For the same Omni-MATH task, SmolAgents’ Manager collects feedback without two-way communication, and the initial error persists; its trajectory is shown in Figure 5.The contrast illustrates the different interaction patterns of the Chair and a standard HMW manager.Source: §6.2 · p. 8, Figure 5 · p. 8

6.3 Error Analysis

  • In a GPQA failure, the Web Agent gives an incorrect symmetry claim and the Chair selects it because it cannot independently verify the retrieval output.The case illustrates how assigning specialized tools exclusively to Workers can let retrieval errors propagate without cross-validation.Source: Table 6 · p. 7, §6.3 · p. 8

7 Conclusion

BusMA provides a shared communication channel for agents to address peers using four intents, with Workers and a monitoring Chair coordinated through Bus modules. Experiments across 13 tasks show stronger performance than state-of-the-art HMW and RMP methods.

7 Conclusion

  • BusMA lets agents directly address peers through a shared channel using discussion, challenge, guidance, and requests for explanation, with registration, routing, and shared-memory management supporting Worker–Chair coordination.The paper plans to study confidence estimation and scalability on tasks involving more agents and longer reasoning chains.Source: §7 · p. 8

Limitations

This section outlines limits on BusMA arising from the underlying model, tool evidence, coordination at scale, and tasks without verifiable intermediate steps. It also documents tool and agent configurations and baseline-model details for clarity and reproducibility.

Limitations

  • BusMA does not improve an LLM’s underlying capabilities, and extra exchanges can add noise when perception is the bottleneck.Incomplete or incorrect tool evidence can mislead Workers and bias the Chair; unfamiliar tools can make outputs hard to verify, contributing to weaker GPQA results. Confidence-aware arbitration and conflict detection are proposed as possible remedies.Source: Limitations · p. 9
  • Experiments typically coordinate only 2–3 agents, while larger groups can cause bus contention as shared memory grows linearly with routed messages and the Chair synthesizes longer histories.Blocking replies, arbitration protocols, and priority scheduling are beyond this work’s scope and proposed for future work; larger-scale benchmarks are also not yet required by widely used multi-agent benchmarks.Source: Limitations · p. 9

B Single-Agent Baseline Comparison

  • BusMA outperforms a single agent on 10 of 12 tasks, with an average improvement of 5.8 percentage points.Gains include Game of 24 (+18.5%), HotpotQA (+13.5%), and MathVista (+13.0%); visual-task results include AlgoPuzzleVQA (−0.5%) and PuzzleVQA (−3.0%).Source: Table 7 · p. 19, B Single-Agent Baseline Comparison · p. 19

C.4.2 Tool Setup

  • GoogleSearchTool retrieves search-result titles, URLs, and snippets, while CodeExecution runs generated code in a temporary directory and returns results.GoogleSearchTool uses the Google Custom Search API; CodeExecution runs code generated by the Code Agent.Source: §C.4.2 · p. 20

C.5.1 Agent Setup

  • The agent setup assigns specialized roles and iteration limits to the Chair, Browser, File, and Code agents.The Chair has no tools and a 10-iteration cap; Browser uses OpenDeepResearch tools with a 20-iteration cap, while File and Code each have a 12-iteration cap. The Code Agent receives execution results at the next iteration.Source: §C.5.1 · p. 20

D Baselines Details

D.1 Diversity Benchmarks

The appendix describes comparison-system configurations for diversity benchmarks and GAIA, including multi-agent frameworks, a single-agent function-calling baseline, and specialized agentic systems.

D.1 Diversity Benchmarks

  • Diversity Benchmarks include OctoTools, SmolAgents, LangGraph, and AutoGen as multi-agent comparison systems.The configurations use planner/executor or manager/supervisor/router roles alongside specialized agents and tools.Source: §D.1 · pp. 20–21

D.2 GAIA

  • Gemini FunctionCalling is a single Gemini-2.5-Flash or Pro invocation with registered functions for search, code execution, and file execution.For image tasks, Gemini uses native image analysis; output is capped at 8,192 tokens.Source: §D.2 · p. 21
  • GAIA comparisons include MagenticOne and OpenDeepResearch, using orchestrator or manager-worker architectures with configured step limits.MagenticOne is capped at 120 steps; OpenDeepResearch sets limits of 12 for its Manager and 20 for its Browser Agent.Source: §D.2 · p. 21

E Qualitative Analysis

The qualitative analysis contrasts BusMA and SmolAgents on the same Omni-MATH problem, showing how their communication trajectories handle an initially erroneous derivation.

Qualitative Analysis

  • Both trajectories address the same Omni-MATH probability problem involving a polygon center and a triangle formed by three randomly chosen vertices.Source: E Qualitative Analysis · p. 21

BusMA Communication Trajectory

  • BusMA’s Chair requests explanation, gives validation guidance, and exchanges challenges with the Code Agent, prompting iterative checks and mutual correction.The agents eventually identify a formula’s limitation at n=1, which the Chair includes as an edge-case caveat in the final answer.Source: E Qualitative Analysis · p. 22

SmolAgents Communication Trajectory

  • SmolAgents’ Manager collects the Code Agent’s result without two-way communication, so the initial error persists in the final output.Source: E Qualitative Analysis · p. 22

F Error Analysis

A GPQA retrieval failure shows how incorrect tool evidence can pass through BusMA’s shared memory to the final answer when the Chair cannot independently verify it.

F Error Analysis

  • The Web Agent incorrectly identified triisopropyl borate as having C3h symmetry, and the Chair selected it because its claim was more confident than the speculative alternatives.Triisopropyl borate belongs to C3 and lacks the horizontal mirror plane required for C3h symmetry.Source: F Error Analysis · p. 22
  • Because the Chair had no web-search access, it could not verify the Web Agent’s retrieval and the error propagated through shared memory into the final answer.The Chair initiated collaboration by requesting symmetry information for each molecule from the Web Agent.Source: F Error Analysis · p. 22
  • The case suggests confidence-aware arbitration with conflict detection could help address errors that other agents cannot independently verify.Unlike qualitative analysis where the Chair could challenge or request an explanation, factual retrieval errors are difficult to detect without independent verification.Source: F Error Analysis · p. 22

G CLEVR-Math Over-Specialization Analysis

A CLEVR-Math trajectory illustrates that delegating trivial computation to a specialized agent can add coordination overhead without improving correctness, so role division should be selective.

G CLEVR-Math Over-Specialization Analysis

  • The CLEVR-Math example asks how many balls remain after subtracting cyan balls, with the Code Agent invoked for the trivial subtraction.Table 9 presents the trajectory for this instance.Source: G CLEVR-Math Over-Specialization… · p. 23, Table 9 · p. 23
  • Delegating trivial mathematical operations to a specialized tool-agent increases inter-agent messages and tool invocations without improving correctness.The analysis notes that fine-grained role division can suppress an LLM’s inherent reasoning and introduce unnecessary communication overhead.Source: G CLEVR-Math Over-Specialization… · pp. 22–23
  • The decision to delegate should depend on task complexity and whether intermediate steps are verifiable, rather than default role assignment..Source: G CLEVR-Math Over-Specialization… · p. 23

Related papers

Loading 2609.15054v2…