Toolformer

Toolformer started with a 2023 Meta paper, 'Toolformer: Language Models Can Teach Themselves to Use Tools'. The core idea is straightforward: instead of a human prompting 'use a calculator here', the LLM is trained to insert candidate tool calls into its own training data and learn from whether those calls actually improved the answer.

Academically the broader trend is usually called Tool Learning. The model internalises tool use as a learned capability, and interface standards like function calling and MCP sit on top of that. Toolformer is less about 'how do you call a tool' and more about 'when and why does the model decide to call one'.

Marketers rarely touch this directly, but it is useful background for why recent models reach for tools so naturally. As tool use becomes a learned capability, the visibility question stretches past 'is the API well-built' into 'does the model naturally reach for our API at all'.