What they did
The authors developed SMITH, a reinforcement-learning framework in which each rollout asks a single language-model policy either to construct a tool from a few examples or to invoke a previously created tool on a held-out question. The training pool covered 13 procedural reasoning tasks with exact, automatically checked answers.
SMITH assigns separate rewards for producing a valid tool schema, writing correct executable code, and reaching the correct final answer. This is intended to give the model distinct learning signals when failures arise from the interface, implementation, or tool use.
Key findings
- A SMITH-trained 4B Qwen3 reached 79.8% macro-average accuracy on held-out tasks, outperforming every evaluated comparison method, including an untrained Qwen3-30B-A3B tool writer.
- The model scored 40.4% on TabMWP-Hard and 42.6% on the out-of-domain GQA benchmark despite receiving no tabular or visual training data.
- On GQA, SMITH improved by 7.6 percentage points over the strongest inference-time tool-creation baseline using the same backbone.
- Tools produced by the trained 4B models also improved task performance when used by LFM-2.5-350M and Qwen3-30B-A3B, suggesting that some learned tools transfer across model sizes and families.
Why it matters
Most tool-creation systems ask a frozen model to invent an API at inference time, without teaching it to design interfaces that it can reliably call. These results indicate that jointly optimizing tool construction and invocation can make a relatively small model more effective than larger, untrained tool writers while also producing tools usable by other models.
Caveats
The main training evidence comes from 13 procedural tasks with exact verifiers, a setting that provides unusually clear reward signals. The reported transfer results cover specific tabular and visual benchmarks; they do not establish reliability on open-ended software development, real-world APIs, ambiguous tasks, or tools with security and side-effect risks.