• Sources: arXiv 2608.24571
  • Summary: The paper presents SMITH, for Schema-grounded Multi-task Iterative Tool Honing, which trains tool creation and tool use in a single policy rather than prompting a frozen model at inference time. Training uses separate reward axes for schema, code, and outcome. The authors report a 4B Qwen3 reaching 79.8 macro-average accuracy against an untrained 30B-A3B tool-writer.
  • Why it matters: Existing tool-creation systems never tell the writer that the schemas it emits are schemas it will have to invoke, and the reported gains come from closing that loop.