SMITH Trains Tool Creation and Tool Use in One Policy; a 4B Qwen3 Beats an Untrained 30B Tool-Writer at 79.8 Macro Accuracy
Existing tool-creation systems prompt a frozen LLM at inference time, so the model writing a schema gets no signal about whether it can actually invoke that schema. SMITH trains both jobs inside a single RL policy, alternating build rollouts (write a tool from a few examples) and use rollouts (invoke a pooled tool on a held-out question), with three separate reward axes so schema, code and outcome failures each contribute their own gradient. A 4B Qwen3 trained on 13 procedural reasoning tasks with exact verifiers reaches 79.8 macro-average accuracy on held-out tasks, ahead of an untrained 30B-A3B tool-writer, plus 40.4 on TabMWP-Hard and 42.6 on out-of-domain GQA (+7.6) with no visual or tabular training data. Tools written by the 4B models also lifted LFM-2.5-350M and Qwen3-30B-A3B.
↳ Follow the thread