Reproduce “MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems” 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

Rui Ye, Shuo Tang, Rui Ge, Yaxin Du, Zhenfei Yin, Siheng Chen, Jing Shao

Shanghai Jiao Tong University and Shanghai AI Laboratory researchers introduce MAS-GPT, a framework that enables automatic generation of task-specific Multi-Agent Systems through a single LLM inference, demonstrating superior performance across 9 benchmarks while reducing computational costs compared to existing approaches that require multiple LLM calls or manual design.

Key insights

  • Multi-agent systems can be effectively represented as executable code with agent prompts as variables and interactions as string concatenations
  • Inter-consistency (similar queries get similar MAS) and intra-consistency (query-MAS alignment) are crucial for training effective MAS generators
  • Medium-sized open-source LLMs can be fine-tuned to generate sophisticated multi-agent systems with a single inference

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.