Research map / Multi-agent systems and games / LLM agents
AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems
Read the paperCheck your idea against 108,967 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 AgentNet, a decentralized RAG-based multi-agent framework where agents route tasks over an evolving DAG-structured graph and refine skills via retrieval memories, evaluated on MATH, BBH and API-Bank against single- and multi-agent baselines.
Problem
Centralized multi-agent LLM systems suffer from scalability bottlenecks, single points of failure, rigid static roles, and privacy barriers to cross-organizational collaboration.
Approach
Removes the central orchestrator: each agent has a router and executor with fixed-size retrieval memories of local step fragments; agents forward, split, or execute tasks along a DAG routing path, and edge weights are updated and pruned based on collaborative success.
Key results
- On BBH with DeepSeek-V3, AgentNet reaches 94.00% accuracy versus 92.00% for the best single-agent... (ReAct 88.00%, GPTSwarm 90.00%) and 92.86% on MATH vs 92.14% for MetaGPT.
- With GPT-4o-mini, AgentNet ties GPTSwarm at 86.00% on BBH and reaches 29.00% on API-Bank, above all baselines except GPTSwarm-free rows.
- With Qwen-turbo, AgentNet scores 92.00% on BBH versus 75.00% for the next best multi-agent method (GPTSwarm), but trails AFlow on MATH (81.43 vs 82.14).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AgentNet instead of MetaGPT | DeepSeek-V3 multi-agent system · MATH test | accuracy: 92.14 -> 92.86 | Helped |
| Use AgentNet instead of Synapse | DeepSeek-V3 · BBH test | accuracy: 92.00 -> 94.00 | Helped |
| Use AgentNet instead of React | DeepSeek-V3 · API-Bank test | accuracy: 29.00 -> 30.00 | Helped |
| Use AgentNet instead of AFLOW | GPT-4o-mini · MATH test | accuracy: 85.00 -> 85.00 | No clear effect |
| Use AgentNet instead of GPTSwarm | GPT-4o-mini · BBH test | accuracy: 86.00 -> 86.00 | No clear effect |
| Use AgentNet instead of React | GPT-4o-mini · API-Bank test | accuracy: 24.00 -> 29.00 | Helped |
| Use AgentNet instead of AFLOW | Qwen-turbo · MATH test | accuracy: 82.14 -> 81.43 | Hurt |
| Use AgentNet instead of GPTSwarm | Qwen-turbo · BBH test | accuracy: 75.00 -> 92.00 | Helped |
| Use AgentNet instead of GPTSwarm | Qwen-turbo · API-Bank test | accuracy: 30.00 -> 32.00 | Helped |
| Add skill heterogeneity to a homogeneous team | AgentNet, 3-agent BBH team · BBH | accuracy: 0.86 -> 0.84 | Hurt |
| Add LLM heterogeneity to a homogeneous team | AgentNet, 3-agent BBH team · BBH | accuracy: 0.86 -> 0.81 | Hurt |
| Add skill heterogeneity to a homogeneous team | AgentNet, 5-agent BBH team · BBH | accuracy: 0.79 -> 0.86 | Helped |
Surprising findings
- AgentNet's MATH accuracy with Qwen-turbo (81.43%) is slightly below AFlow (82.14%), and on API-Bank gains are small (e.g., 30.00 vs 29.00 with DeepSeek-V3 against single-agent baselines).
- In the heterogeneity study, the fully homogeneous setting performs best with 3 agents, while heterogeneous settings help only with 5 agents.
Limitations
- Router selects from a small predefined candidate pool; scalability to hundreds or thousands of agents is untested.
- Heterogeneous agent environments remain an open challenge, with mixed results on small teams.
- Experiments use small custom train/test splits (e.g., 100 test problems per benchmark) with a single run and no variance reporting.
Experimental setup
Datasets: MATH BBH API-Bank
Benchmarks: MATH BBH API-Bank
Metrics: accuracy
Baselines: Direct ReAct Synapse Self-Consistency Self-Refinement MorphAgent MetaGPT AFlow GPTSwarm
Builds on: ReAct RAG MorphAgent AFlow GPTSwarm
Closest papers by meaning
- Augmented Runtime Collaboration for Self-Organizing Multi-Agent Systems: A Hybrid Bi-Criteria Routing Approach (AAAI 2026 · similarity 0.90)
- Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs (NeurIPS 2026 · similarity 0.90)
- Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.89)
- TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (ICML 2026 · similarity 0.89)
- Multi-Agent Teams Hold Experts Back (ICML 2026 · similarity 0.88)
- Emergent Coordination in Multi-Agent Language Models (ICLR 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.