MidTool: Mid-training Data Synthesis for Agentic Tool Use
The paper argues tool use should be trained before post-training, not patched in afterward.
MidTool builds an open corpus pipeline from web, PDF, code, real tool APIs, MCP skills, and document-grounded workflows. The training target is practical agent behavior: spotting what tools can do, filling arguments from context, chaining calls, and recovering when information is missing. The authors report gains for Qwen3-4B-Base and Qwen3-8B-Base after MidTool-Mix mid-training, followed by SFT and reinforcement learning. Benchmarks named in the abstract are BFCL, tau2-Bench, and MCP Universe. ArXiv · AI/CL/LG's note
MidTool builds an open corpus pipeline from web, PDF, code, real tool APIs, MCP skills, and document-grounded workflows. The training target is practical agent behavior: spotting what tools can do, filling arguments from context, chaining calls, and recovering when information is missing. The authors report gains for Qwen3-4B-Base and Qwen3-8B-Base after MidTool-Mix mid-training, followed by SFT and reinforcement learning. Benchmarks named in the abstract are BFCL, tau2-Bench, and MCP Universe. ArXiv · AI/CL/LG's note
score 5