Tool use can be learned from the language model’s own prediction errors. The useful signal is whether an API response makes subsequent text easier to predict, allowing ordinary text to supply supervision for choosing tools and arguments.
Key Findings
- A few demonstrations seed candidate calls; execution and loss-based filtering create the fine-tuning data without extensive manual annotation.
- On ASDiv, the GPT-J-based model scores 40.4 versus 14.8 with its APIs disabled, showing the contribution of calculator access.
- Experiments cover question answering, search, arithmetic, translation, and dates while retaining general language modeling ability.
Who Should Read It
Great fit if you study how to generate tool-use training data. Look elsewhere if you need chained calls or interactive search: the method does not teach either, and useful examples for some tools are costly to obtain.