BusMA: A Bus Communication Substrate for Multi-Agent Systems
Yanwen Peng, Delvin Ce Zhang, Xi Wang, Nikolaos Aletras
cs.AI
Quick verdict
BusMA lets agents address peers on a shared bus with four intents; wins on 13 benchmarks look broad, but with no significance tests small margins stay unproven.
Summary
BusMA replaces hub-mediated communication (hierarchical manager-worker, router-based message passing) in LLM multi-agent systems with a shared 'bus' substrate inspired by computer architectures. Workers with tools keep private local memory and post JSON messages to specific peers using four intents (discussion, challenge, guidance, request for explanation), while a tool-less Chair agent monitors shared memory and coordinates convergence. Evaluation covers 12 benchmarks (200 samples each) plus GAIA (165 questions) with DeepSeek-V3 and Gemini-2.5-Flash backbones against OctoTools, SmolAgents, LangGraph, and AutoGen. BusMA attains the best averages (68.6% with DeepSeek-V3, +4.8 over OctoTools; 76.3% with Gemini-2.5-Flash, +2.5) and 48.5% on GAIA (+4.3 over OpenDeepResearch). Ablations show swapping in SmolAgents workers costs at most 1.5%, attributing gains mainly to the communication structure, while replacing the Chair with a manager costs 4.5-7.0 points on two of three tasks. However, per-dataset deltas range from -4.5 (GPQA, DeepSeek-V3) to +13.5, no significance tests or multi-seed variance are reported despite n=200, some baseline scores (LangGraph 30.5% on HotpotQA) suggest configuration issues, and token/latency costs are unreported. The authors candidly flag error propagation through shared memory and the 2-3-agent scale limit. Anonymous code is available.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Worker-initiated, intent-tagged peer-to-peer messaging over a shared bus outperforms state-of-the-art HMW and RMP frameworks across 13 benchmarks (average +4.8 with DeepSeek-V3, +2.5 with Gemini-2.5-Flash).
- The communication structure, not the worker implementation, accounts for most of the gains (replacing BusMA Workers with SmolAgents Workers costs at most 1.5%).
- On GAIA, BusMA reaches 48.5% overall, +4.3 over OpenDeepResearch, while reusing the same browser agent.
- Gains concentrate on tasks with verifiable intermediate steps (math, retrieval), while perception-bounded visual tasks gain little or regress.
- Worker heterogeneity creates channels for unverifiable claims that propagate errors through shared memory when tools are exclusively held.
Strengths
Broad, protocol-matched evaluation
13 benchmarks (12 diversity sets at 200 samples plus GAIA-165), two backbones, four baselines; visual tool calls uniformly routed to Gemini-2.0-Flash under DeepSeek-V3, and GAIA reuses OpenDeepResearch's browser agent via adapters, so the Table 4 gap plausibly reflects coordination, not tooling.
Well-conceived ablations that isolate the communication layer
Table 5: swapping BusMA Workers for SmolAgents implementations moves results by at most 1.5% (and beats SmolAgents by 4.0-10.0%), while replacing the Chair with an HMW manager costs 4.5-7.0 points on PuzzleVQA/MathVista; the single-agent baseline (Appendix B, +5.9 average) adds a third comparison axis.
Transparent failure and cost-of-specialization analyses
Reproducibility hygiene above area norm
Weaknesses
No statistical inference anywhere
'Consistently outperforms' overstates Table 3
Baseline configuration and budget asymmetries undiscussed
Internal inconsistency in ablation narration
The bus metaphor's load-bearing features are untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
BusMA (arXiv 2609.15054, cs.AI, Sheffield) targets the communication layer of LLM multi-agent systems. The authors argue that hub-mediated topologies - Hierarchical Manager-Worker (HMW: AgentVerse, ChatDev, OctoTools) and Router-based Message Passing (RMP: AutoGen) - leave workers as passive executors and turn the hub into a cognitive bottleneck that propagates errors. Inspired by hardware buses and Oki et al.'s Information Bus, BusMA provides agent registration, per-agent FIFO routing queues, and append-only shared memory. Workers run a ReAct-style loop with private local memory and post JSON messages with one of four intents (discussion, challenge, guidance, request for explanation); a tool-less Chair ('primus inter pares') monitors shared memory, issues follow-ups, and synthesizes the final answer. Central claims: (1) BusMA beats SOTA HMW/RMP frameworks on 13 tasks with two backbones; (2) communication structure, not worker implementation, drives gains; (3) gains concentrate where intermediate steps are verifiable.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is BusMA and how does it differ from AutoGen or manager-worker frameworks?
BusMA is a communication layer for LLM multi-agent systems in which every agent can directly message any other agent through a shared 'bus' that handles registration, routing, and shared memory, using four intents: discussion, challenge, guidance, and request for explanation. A tool-less Chair agent coordinates and synthesizes but does not route every message. In manager-worker hierarchies and router-based systems like AutoGen, exchanges pass through a central hub, so workers cannot directly consult specific peers and the hub can propagate errors.
How well does BusMA perform on benchmarks?
It achieves the best average accuracy on 12 diverse benchmarks with two LLM backbones (68.6% with DeepSeek-V3, beating the best baseline by 4.8 points; 76.3% with Gemini-2.5-Flash, +2.5) and scores 48.5% on GAIA versus 44.2% for OpenDeepResearch. It loses several individual visual/knowledge datasets, and the paper reports no statistical significance tests, so per-task wins of 1-3 points should be read cautiously.
Is BusMA code publicly available?
Yes - the paper links an anonymous repository (anonymous.4open.science/r/Bus-MA-370E) and provides all agent prompts, JSON communication protocols, and configuration details in its appendices, though running it requires paid LLM APIs and a Google Custom Search key.
Does BusMA beat a single agent?
With Gemini-2.5-Flash, yes on 10 of 12 tasks, by an average of +5.9 points (Appendix B, Table 7), with the largest gains on verifiable-step tasks such as Game of 24 (+18.5) and HotpotQA (+13.5); on some visual tasks it is slightly worse (PuzzleVQA -3.0, AlgoPuzzleVQA -0.5).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are present only as caption text in the extraction; qualitative judgments rely on Tables 6, 8, and 9 instead of the rendered diagrams.
- The extraction contains minor OCR artifacts (e.g., 'Shared Memery', truncated citation years); all cited numbers were cross-checked against the result tables.
- The anonymous code repository was not inspected; reproducibility rating rests on the paper's description of Appendices A-D.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.