Summary

  • Researchers from the Soochow University of China have created a new system called Chain-of-Tools (CoTools) that enhances large language models’ (LLMs) use of external tools.
  • While current LLMs are very good at text generation and understanding, they must use external resources and applications to perform many tasks.
  • This requires them to be trained on, or tuned to use, these specific tools, which limits their flexibility and can impact their core abilities.
  • The CoTools system combines aspects of fine-tuning and in-context learning while keeping the core LLM frozen, meaning its original weight and reasoning capabilities are left untouched.
  • Three key components are used: a Tool Judge, a Tool Retriever and a Tool Caller.
  • The Tool Judge decides whether a tool is needed at a certain point, the Retriever selects the best one, and the Caller uses in-context learning (ICL) prompts to fill in the tool’s parameters, allowing it to select new or unused tools efficiently.
  • The team believes it can drive the development of more flexible, powerful LLM-powered agents that can more easily adapt to new tools and APIs.

By Ben Dickson

Original Article