Toolformer: Language Models Can Teach Themselves…

Toolformer asks how a language model can learn when and how to call external tools without a large hand-labeled tool-use dataset. Read it to see how the model proposes calls inside ordinary text, executes them, and keeps only those whose returned result lowers continuation loss enough to become a training example.

Reading focus: How a few demonstrations turn the base language model into a permissive annotator that proposes API-call positions and arguments. Why Toolformer compares future-token loss with and without each executed result, and what that threshold does not certify. How retained calls become fine-tuning data, why sufficiently capable base models benefit most, and where single-call learning stops short of a complete agent.

NeurIPS 2023. Schick et al.. 40 min read, easy difficulty.