Reproduce “Learning to Use Tools via Cooperative and Interactive Agents” 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
Zhengliang Shi, Shen Gao, Xiuyi Chen, Yue Feng, Lingyong Yan, Haibo Shi, Dawei Yin, Pengjie Ren and 2 more
The ConAgents framework enhances Large Language Models' (LLMs) ability to use external tools by employing a cooperative and interactive multi-agent system. This approach achieves up to 14% higher success rates on tool-use benchmarks compared to single-agent methods and effectively enables specialized tool-use capabilities in open-source LLMs through distillation.
Key insights
- Decomposing complex tool-use workflows into specialized roles for distinct LLM agents, combined with dynamic inter-agent communication, leads to more robust and accurate task completion.
- A dedicated Review Agent providing real-time, constructive feedback on both planning and execution is crucial for effectively calibrating errors and improving an agent system's self-correction capabilities.
- Specialized Action Distillation (SPAN) effectively transfers sophisticated tool-use knowledge from powerful proprietary models to smaller, open-source LLMs, making advanced agent capabilities more accessible.
- A well-designed multi-agent architecture can achieve higher performance and sometimes even be more token-efficient than single-agent methods by reducing exploratory "hallucination" and redundant attempts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.