Reproduce “MetaGen: Self-Evolving Roles and Topologies for Multi-Agent LLM Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yimeng Wang, Jiaxing Zhao, Hongbin Xie, Hexing Ma, Yuzhen Lei, Shuangxue Liu, Xuan Song, Zichen Zhang and 1 more
Researchers from Jilin University, SUSTech, and Peking University developed MetaGen, a training-free framework that allows multi-agent LLM systems to dynamically evolve both agent roles and their collaboration topology during inference. This framework achieved state-of-the-art accuracy with significantly reduced token costs and enhanced adaptability across various complex reasoning tasks.
Key insights
- Simultaneous and dynamic adaptation of agent roles and collaboration topology during inference is crucial for building flexible and robust multi-agent LLM systems.
- A training-free approach, leveraging an Architect Agent and lightweight feedback, can achieve superior performance and cost efficiency without modifying base LLM weights.
- Combining a minimal, stable task backbone with a dynamically evolving set of query-conditioned roles allows for both foundational structure and high adaptability.
- Accumulating verified roles and updating selection priors across tasks significantly enhances system robustness and improves performance on subsequent, similar instances.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.