The Pulse
SMITH Trains a 4B Model to Build and Use Its Own Tools
Appier says its SMITH research paper has been accepted at NeurIPS 2026. The system jointly trains a 4-billion-parameter Qwen3 model to create reusable tools and call them, with tests showing transfer to smaller and larger models.
AI.info Team · Written with AI ·
Appier said on September 30 that its SMITH research paper had been accepted at NeurIPS 2026, advancing a claim that runs against a common approach to AI tools: assigning a large model to write a tool and a separate model to use it. SMITH instead trains one model to do both, using the results of tool calls to improve how it builds tools. The paper appeared on arXiv in August and reports that a 4-billion-parameter Qwen3 model outperformed larger tool-writing baselines on held-out procedural reasoning tasks.
One model writes the tool and has to use it
Existing tool-creation methods often generate a function at inference time, without training the writer on whether another model can understand or call it reliably. SMITH, short for Schema-grounded Multi-task Iterative Tool Honing, combines tool creation and use in a single reinforcement-learning policy. The model creates a Python function and a matching JSON schema; when it later uses the tool, it sees the schema rather than the underlying code.
“We propose SMITH (Schema-grounded Multi-task Iterative Tool Honing), a reinforcement learning framework that jointly trains tool creation and tool use inside a single policy,” write Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, Shao-Hua Sun and Hung-yi Lee, the paper’s authors. Its training alternates between building tools from examples and using tools to answer questions. Separate rewards assess whether the schema and code are valid and whether the tool produces correct answers.
Four examples, then 16 harder questions
For a build task, the model receives four question-and-answer examples and produces a function and its schema. Researchers test that tool on 16 held-out questions drawn from a harder difficulty band, then add tools that pass to a shared pool for later tasks. The setup is designed to reward a reusable procedure rather than a tool that simply matches the examples.
The main experiment fine-tunes Qwen3-4B-Instruct on 13 procedural task categories from Reasoning-Gym, including arithmetic, algorithms, algebra, games and logic. The authors also evaluate on 10 Reasoning-Gym categories withheld from training. The paper’s abstract reports 79.8% macro-average accuracy on those held-out tasks; its results table reports 79.9%, a small inconsistency within the paper.
Tools transfer across model sizes
The authors compare SMITH with inference-time tool-writing systems and larger-model baselines. On the held-out Reasoning-Gym tasks, the paper’s table gives SMITH 79.9% accuracy, against 74.1% for a Qwen3-30B-A3B model used as a tool-writer. The researchers also report an average of 100 output tokens for SMITH, compared with 3,206 for standard chain-of-thought reasoning on the same evaluation.
Tools created by the trained 4B model also helped other models. In the paper’s tests, an LFM2.5 model with 350 million parameters rose from 11.6% to 42.9% held-out accuracy when using SMITH’s tools, slightly above the 41.5% result with tools from the untrained 30B-A3B writer. A Qwen3-30B-A3B consumer also scored better with SMITH-created tools than when it wrote tools for itself.
Benchmark gains are not a deployment test
SMITH reached 40.4% on TabMWP-Hard, a strengthened tabular-reasoning benchmark, and 42.6% on the visual-question-answering benchmark GQA, neither of which appeared in training. On GQA, the paper ranks SMITH second, behind a baseline distilled from GPT-4.1; the authors explicitly say they do not claim parity on visual perception. Those results indicate transfer beyond the training tasks, but they do not establish performance in live business workflows.
Appier’s announcement framed the work as a route to agents sharing proven tools. The experiments support the narrower claim that a jointly trained model can create tools that other tested models reuse on benchmark problems. The paper’s central result is a measured one: on its held-out procedural tasks, a 4B tool-maker beat the larger tool-writing baseline the authors evaluated.