Reproduce “ToolGen: Unified Tool Retrieval and Calling via Generation” 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
Renxi Wang, Xudong Han, Lei Ji, Shu Wang, Timothy Baldwin, Haonan Li
ToolGen transforms how large language models interact with external tools by unifying tool retrieval and calling into a single generative task. This approach enables LLMs to autonomously select and use tens of thousands of real-world tools with high efficiency, significantly reducing hallucination of tool names.
Key insights
- Tool knowledge can be directly embedded into an LLM's parameters through tool virtualization, representing each tool as a unique, new vocabulary token.
- Unifying tool retrieval and calling into a single generative process eliminates the need for external retrieval modules and reduces context window burden.
- Constrained beam search during inference effectively prevents the generation of non-existent or hallucinated tool names, enhancing reliability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.