SMITH: Joint Tool Creation & Use

The efficacy of tool-augmented language models has been historically constrained by the finite set of APIs humans have developed. Existing approaches to tool creation often decouple the model that writes a tool from the one that uses it, leading to a lack of direct feedback on schema usability. This separation creates a significant gap, as the tool-writing model receives no signal regarding the schemas it produces are indeed invocable by its counterpart.

Bridging the Tool Creation-Use Chasm

A novel approach, SMITH (Schema-grounded Multi-task Iterative Tool Honing), introduces a reinforcement learning framework designed to tackle this challenge head-on. SMITH uniquely enables the joint training of both tool creation and tool use within a single policy. This integrated approach means that each training rollout is strategically segmented into either a build task, where the model learns to write a tool from provided examples, or a use task, where it invokes a pooled tool to answer a held-out question.

Granular Feedback for Robust Tooling

The framework's strength lies in its sophisticated reward system. SMITH employs three distinct reward axes to independently assess failures in schema generation, code correctness, and task outcome. This granular feedback mechanism ensures that each specific failure mode contributes its own gradient signal during training, facilitating more targeted and effective learning. This contrasts with prior methods that might offer a more generalized failure signal.

The impact of this integrated training is substantial. A 4B Qwen model, trained with the SMITH reinforcement learning framework on 13 procedural reasoning tasks featuring exact verifiers, achieved a 79.8% macro-average accuracy on held-out tasks. This performance surpasses all other evaluated methods and notably outperforms an untrained 30B-A3B tool-writer. Furthermore, the SMITH reinforcement learning framework demonstrated impressive generalization capabilities, reaching 40.4% accuracy on TabMWP-Hard and 42.6% on out-of-domain GQA, an improvement of +7.6% over the best same-backbone inference-time baseline, without any specialized visual or tabular training data.

Beyond its own performance, tools generated by the 4B SMITH models significantly boosted the capabilities of larger models, including LFM-2.5-350M and Qwen3-30B-A3B, on the same reasoning tasks. This highlights the SMITH reinforcement learning framework's potential to enhance existing LLM ecosystems by producing high-quality, usable tools.