Reproduce “Optima: Optimizing Effectiveness and Efficiency for LLM-Based Multi-Agent System” 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
Weize Chen, Jiarui Yuan, Chen Qian, Cheng Yang, Zhiyuan Liu, Maosong Sun
Researchers at Tsinghua University developed OPTIMA, an iterative training framework for Large Language Model (LLM)-based multi-agent systems, simultaneously optimizing communication efficiency and task effectiveness. The framework yields up to a 2.8x F1 performance gain on multi-hop reasoning tasks while reducing token usage by over 90%, significantly enhancing inference-time scaling and generalization.
Key insights
- Explicitly training LLMs within a multi-agent context with an efficiency-aware reward function drastically improves both communication efficiency and task effectiveness.
- A balanced reward function, combining task performance, token reduction, and language modeling loss, is crucial for developing concise, interpretable, and effective inter-agent communication.
- OPTIMA-trained agents acquire robust, transferable skills for efficient information exchange and collaborative reasoning, demonstrating strong generalization across similar tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.