Toolformer: Language Models Can Teach Themselves to Use Tools

Language models are oddly bad at things that simple programs do perfectly: dividing two numbers, looking up a fact that changed last year, knowing today's date. The obvious fix is to let the model call tools. The question in early 2023 was how to teach a model when to call which tool, with what input, and how to use the answer, without a large hand-labeled dataset of tool use.

Timo Schick, Thomas Scialom and colleagues at Meta AI Research answer with a self-supervised trick. Write a short prompt with a few examples per tool, then let the model annotate a large slice of web text with candidate calls, such as [Calculator(400/1400) → 0.29] placed just before "29%". Execute each candidate call. Then keep a call only if putting the call and its result in front of the text makes the model better at predicting the next few words than having no call, or the call without its result. Fine-tune the model on the text with the surviving calls. Because the text is ordinary pretraining-style data, the model keeps its general abilities.

The tools were a question-answering model, a Wikipedia search engine, a calculator, a calendar and a translation model. The base model, GPT-J with 6.7 billion parameters, became much stronger zero-shot on fact completion, arithmetic word problems and date questions, often beating GPT-3 with 175 billion parameters. It still lagged on open question answering and could neither chain tools nor search interactively.

Modern models arrive already trained to call functions, so few teams will reproduce Toolformer. The ideas in it still decide whether a customer's tool-using system works.

Tools fix what models are structurally bad at. Exact arithmetic, current facts and dates are the cases where tools gave the largest gains. If a customer workflow depends on any of them, route it to a tool rather than hoping the model computes or remembers.

"Did the tool help?" is measurable. Toolformer's filter asks whether the tool's result made the following text more predictable. In evaluation you can ask the same question: does the call change the outcome, or is it ceremony that adds latency?

Placement matters. A call helps most right before its result is needed. In agents, fetch information close to where it is used, rather than front-loading everything into the context.

Small models may not benefit. Below roughly 775M parameters, the models did no better with tools than without. When a customer insists on a very small model, test whether tool calls help at all.

Single calls are not enough for real work. Toolformer makes at most one call per output and cannot chain tools. Most business tasks need several calls and decisions between them, which is where agent loops and code actions come in.