MidTool: Mid-training Data Synthesis for Agentic Tool Use
MidTool trains general tool use before post-training, using API, MCP, document, web, PDF, and code data. Qwen3 4B and 8B variants improved across three downstream tool-use benchmarks.
MidTool builds an open mid-training corpus from web, PDF, and code data plus real-world APIs, MCP skills, and document-grounded workflows. It targets tool selection, argument grounding, workflow composition, and incomplete information.
The authors mid-trained **Qwen3-4B-Base** and **Qwen3-8B-Base**, then applied both supervised fine-tuning and reinforcement learning. Builders training agent models should consider tool competence a dedicated training stage rather than only a post-training behavior.
MidTool builds an open mid-training corpus from web, PDF, and code data plus real-world APIs, MCP skills, and document-grounded workflows. It targets tool selection, argument grounding, workflow composition, and incomplete information. The authors mid-trained **Qwen3-4B-Base** and **Qwen3-8B-Base**, then applied both supervised fine-tuning and reinforcement learning. Builders training agent models should consider tool competence a dedicated training stage rather than only a post-training behavior. MidTool-Mix improved results under SFT and RL on **BFCL, tau2-Bench, and MCP Universe**, but the supplied material gives no effect sizes. The evidence is also limited to two sizes from one model family.
MidTool shifts tool competence from an SDK-only or post-training concern into the model’s mid-training curriculum. It supplies open data spanning selection, grounded arguments, workflow composition, and missing information, with improvements under both SFT and RL. The claim remains narrow because effect sizes are absent and evidence covers only Qwen3 at 4B and 8B; it does not establish production reliability.