ToolGen: Unified Tool Retrieval and Calling via Generation
Notes on ToolGen: Unified Tool Retrieval and Calling via Generation by Renxi Wang, Xudong Han, Lei Ji, Shu Wang, Timothy Baldwin, Haonan Li.
A Tool Use paper, first released in October 2024, that represents each tool as a unique token in the LLM's vocabulary.
The usual approach is to put tool descriptions in the prompt, which is limited by context length and needs a separate retrieval step to pick which tools to include. ToolGen instead trains the tool knowledge into the model's parameters, so picking a tool and calling it are just part of next-token prediction. No separate retriever needed.
They test it with over 47,000 tools and report better results for both tool retrieval and completing tasks autonomously.
