Large language models that use tools are limited by the set of human-written APIs. Existing tool-creation systems prompt a frozen model to write tools at inference time, but the model that writes the tool is decoupled from the one that uses it, so there is no feedback loop to improve tool usability. SMITH (Schema-grounded Multi-task Iterative Tool Honing) trains a single policy with reinforcement learning to both create and use tools, alternating between build tasks (write a tool from examples) and use tasks (invoke a tool on a held-out question). Three separate reward axes penalize schema, code, and outcome failures independently, giving each failure mode its own gradient signal.
A 4B Qwen3 model trained with SMITH on 13 procedural reasoning tasks achieved 79.8 macro-average accuracy on held-out tasks, the best among all evaluated methods and ahead of an untrained 30B-A3B tool-writer. On TabMWP-Hard (a hard tabular reasoning benchmark) it reached 40.4, and on out-of-domain GQA (visual reasoning) it reached 42.6, a +7.6 improvement over the best same-backbone inference-time baseline, despite having no visual or tabular training data. Tools written by SMITH-trained 4B models also lifted the performance of LFM-2.5-350M and Qwen3-30B-A3B on the same reasoning tasks.
The significance is that jointly optimizing tool creation and use makes each role inform the other, producing tools that are both correct and easy to invoke. This contrasts with prior work that treats tool writing and tool calling as separate stages, and demonstrates that a single, unified training objective can lead to better performance and cross-model transfer.