<- All papers

Toolformer: Language Models Can Teach Themselves to Use Tools

Explores how language models can learn to call external tools through self-supervised data generation.

Toolformer is useful as a lens for deciding when a workflow should remain a prompt pattern and when it deserves a stable tool interface.

Builder Notes

  • Tool calls need clear affordances, predictable inputs, and cheap failure modes.
  • Self-supervision helps produce training examples, but product workflows still need explicit verification.
  • The most reusable tool interfaces should be documented like APIs, not hidden as prompt tricks.