Research map / Multi-agent systems and games / LLM agents
MPAS: Breaking Sequential Constraints of Multi-Agent Communication Topologies via Individual-Epistemic Message Propagation
Jingxuan Yu, Ju Jia, Simeng Qin, Xiaojun Jia, Siqi Ma, Yihao Huang, Yali Yuan, Guang Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MPAS, a multi-agent communication framework that breaks the sequential topological-order constraint of LLM agent graphs via individual-epistemic message propagation.
Problem
Existing LLM multi-agent systems arrange agent communication as a directed acyclic graph executed one-by-one via topological sort, which imposes sequential constraints on agent communication.
Approach
Replaces sequential topological-order execution of agent graphs with individual-epistemic message propagation, letting agents communicate without a strict one-by-one ordering. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace sequential execution with node-wise parallel message propagation | multi-agent topology · six benchmarks (handcrafted topologies) | accuracy: +0.9% to +6.4% | Improved |
| use MPAS with optimized topology instead of G-Designer | multi-agent system · MMLU | accuracy: 76.9 -> 77.3 | Improved |
| use MPAS optimized topology instead of Agent-Prune | multi-agent system · HumanEval | accuracy: 84.9 -> 87.2 | Improved |
| use parallel MPAS instead of sequential G-Designer at 11 agents | communication · AQuA | time per round (s): 84.6 -> 14.2 | Improved |
| use MPAS instead of G-Designer with 3 agents | communication · AQuA | time per round (s): 7.6 vs 7.8 | No clear effect |
| use MPAS under backdoor attack | multi-agent system · MMLU | accuracy: GPTSwarm 45.8 -> MPAS 72.1 | Improved |
| add attention aggregator (clean) (ablation) | MPAS · MMLU, GSM8K, AQuA, HumanEval | accuracy: 77.3 -> 78.7 (MMLU) | Improved |
| add attention aggregator (clean) (ablation) | MPAS · MultiArith | accuracy: 96.7 -> 95.8 | Hurt |
| add selective aggregator under backdoor (ablation) | MPAS · MMLU backdoor | accuracy: 72.5 -> 77.1 | Improved |
| add attention aggregator under backdoor (ablation) | MPAS · MMLU backdoor | accuracy: 72.5 -> 67.2 | Hurt |
| add selective aggregator in clean setting (ablation) | MPAS · MMLU | accuracy: 77.3 -> 76.6 | Hurt |
| increase communication rounds beyond two on small graphs (ablation) | MPAS · five-agent cycle | performance | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- G-Designer — G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.89)
- MOC: Multi-Order Communication in LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.89)
- Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.89)
- Stochastic Self-Organization in Multi-Agent Systems (ICLR 2026 · similarity 0.89)
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.88)
- Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation (AAAI 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
