The idea in one paragraph.
A model can teach itself when and how to call APIs by self-annotating a corpus with candidate calls and keeping only those that reduce its own loss.
THE MACHINERY
Self-supervised API annotation
Sample candidate calls, execute them, keep those that lower perplexity on the following tokens.
Loss-reduction filter
Use the model's own loss as the verifier for whether a call was worth making.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Trains a model to decide which APIs to call, when to call them, what arguments to pass, and how to incorporate results.
Inspect a source excerpt
a model trained to decide which APIs to call, when to call them, what arguments to pass, and how to…Read the passage in context on arXiv ↗
Achieves substantially improved zero-shot performance across a variety of downstream tasks, often competitive with much larger models, without sacrificing core language modelling ability.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Tools are fixed at training time; the set cannot grow at inference.
- Each call is independent - no composition of tools into procedures.
Why it belongs in this atlas.
The complement to WebGPT: rather than prompting tool use, bake it in. Both stop at the same ceiling - a tool list someone else wrote.