Toolformer: Language Models Can Teach Themselves to Use Tools
ReadA model labels its own training data with API calls that improve its predictions, then learns to make them. Shows how tool use can be trained rather than prompted.
My notes
The filtering criterion is the clever part: keep an API call only if it makes the following tokens easier to predict. Self-supervised usefulness.
Notes
- No human annotation of tool use needed — only a handful of demonstrations per tool.
- Only helps at larger model sizes; small models cannot use the results well.
- Today most tool use is prompted + trained via RL, but the "was this call actually useful?" signal is still the right question for evals.