4B model beats untrained 30B tool-writer

A new training method pushed a 4B Qwen3 model past an untrained 30B-A3B tool-writer, reaching 79.8 accuracy on tasks it never saw in training. Most tool-creation systems prompt a frozen model at inference time, so the model writing a schema gets no signal on whether it works when called. SMITH fixes that by training one reinforcement-learning policy on two kinds of rollouts, building tools and using them, together.

Each rollout type earns three separate reward signals: one for schema correctness, one for code correctness, one for outcome. A failure in any one produces its own gradient, instead of a single blended score that hides which part broke.

The 4B model trained on just 13 procedural reasoning tasks. It then reached 42.6 on out-of-domain GQA, a visual question-answering benchmark, despite training on no visual data at all.

Tools the 4B model wrote also lifted a separate 350M model and a 30B model when handed to them directly. The joint training that made the 4B a better tool-writer also produced tools that generalize across model sizes.

The paper doesn't say whether the gains hold past SMITH's 13 training tasks, or how its reward shaping compares to standard reinforcement learning.