Training LLMs to write tools generalized beyond self use (arxiv.org)

Researchers have introduced a new reinforcement learning framework named SMITH (Schema-grounded Multi-task Iterative Tool Honing) that enables large language models (LLMs) to simultaneously create and utilize tools within a single policy. Traditional tool-augmented models rely on pre-defined APIs, which limits their flexibility. SMITH improves this by training models in two phases—tool creation and tool invocation—allowing them to understand the schemas they generate, ensuring better integration of creation and usage. The framework employs three distinct reward metrics to tackle different types of failures, refining the model's capabilities over time. Significantly, a 4 billion parameter model, Qwen3, trained with SMITH achieved a remarkable 79.8 macro-average accuracy across 13 procedural reasoning tasks, outperforming larger models and traditional methods. This advancement highlights the potential for LLMs to innovate beyond preset APIs, fostering greater adaptability and performance. Additionally, the tools generated by the Qwen3 model boosted the performance of smaller models, indicating a strong transferability of learning and further implications for the development of AI tools that are more robust and capable of tackling complex reasoning tasks without reliance on visual or tabular data.