MidTool: Mid-training Data Synthesis for Agentic Tool Use

TL;DR - MidTool is a data-construction pipeline and open corpus for teaching LLMs general tool use during mid-training. Experiments on Qwen3 base models suggest dedicated mid-training improves agentic tool-use performance beyond relying solely on supervised or reinforcement-learning post-training.

• Synthesizes supervision from real-world APIs, MCP skills, and document-grounded workflows using web, PDF, and code data.
• Targets tool-affordance recognition, context-grounded arguments, multi-step tool-call composition, and recovery from incomplete information.
• MidTool-Mix was used to mid-train Qwen3-4B-Base and Qwen3-8B-Base before follow-up supervised fine-tuning and reinforcement learning.
• It consistently improved downstream results over baselines under both post-training methods on BFCL, tau2-Bench, and MCP Universe.