Research map / Multi-agent systems and games / LLM agents
EvoMAS: Heuristics in the Loop—Evolving Smarter Agentic Workflows
Yangbo Wei, Zhen Huang, Ronghao Xu, Hong Wang, WEI W. XING
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 EvoMAS, an evolutionary framework that automatically constructs and optimizes multi-agent system workflows using a diverse repertoire of heuristics, aiming for efficiency across task complexity levels.
Problem
Constructing efficient multi-agent systems requires labor-intensive manual design, and existing automated methods generate templated agents, use monolithic optimization, and ignore task complexity gradients.
Approach
A biologically-inspired evolutionary framework maintaining a dynamic, diverse repertoire of seven (heuristic-based) components evolved within the loop to produce agentic workflows adapted to task complexity, rather than templated agents or monolithic optimization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add exploration strategies (ablation) | EvoMAS · MATH | accuracy (%): 48.10 -> 60.33 | Improved |
| add exploitation strategies (ablation) | EvoMAS · MBPP | Pass@1 (%): 72.14 -> 86.21 | Improved |
| add the Cyber Creator meta-controller (ablation) | EvoMAS · MATH | accuracy (%): 56.00 -> 60.33 | Improved |
| add the Cyber Creator meta-controller (ablation) | EvoMAS · MATH | cost ($): 3.12 -> 2.54 | Improved |
| add curriculum-guided evolution (ablation) | EvoMAS · MBPP | Pass@1 (%): 83.57 -> 86.21 | Improved |
| add curriculum-guided evolution (ablation) | EvoMAS · MATH | accuracy (%): 57.80 -> 60.33 | Improved |
| add evolution resource library (ablation) | EvoMAS · MBPP | Pass@1 (%): 76.90 -> 86.21 | Improved |
| replace optimizer LLM with Qwen-4B (ablation) | EvoMAS · MBPP | Pass@1 (%): 86.2 -> 74.7 | Hurt |
| replace Judge LLM with Qwen-4B (ablation) | EvoMAS · MBPP | Pass@1 (%): 86.2 -> 84.8 | No clear effect |
| replace Cyber Creator LLM with Qwen-4B (ablation) | EvoMAS · MBPP | Pass@1 (%): 86.2 -> 81.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AgentSquare — AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EvoMAS: Evolutionary Generation of Multi-Agent Systems (ICML 2026 · similarity 0.94)
- Automated Design of Agentic Systems (ICLR 2025 · similarity 0.89)
- AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.88)
- Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 2026 · similarity 0.88)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.88)
- MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.87)
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.
