MidTool: Mid-training Data Synthesis for Agentic Tool Use

Can general tool-use capabilities in large language models be effectively shaped earlier through dedicated mid-training, rather than relying almost entirely on traditional post-training methods like supervised fine-tuning and reinforcement learning?

The authors introduce MidTool, an open corpus construction pipeline that synthesizes agentic mid-training data by combining large-scale web, PDF, and code documents with structured tool APIs and Model Context Protocol (MCP) skills. The pipeline uses a two-branch synthesis strategy: context-grounded trajectory augmentation to extract tool affordances and workflows from unstructured technical documents, and native agentic trajectory synthesis to build executable multi-turn workflows. The resulting 20.3B-token corpus, MidTool-Mix, is used to mid-train Qwen3-4B-Base and Qwen3-8B-Base models, followed by standard supervised fine-tuning and reinforcement learning recipes.

Mid-training models on MidTool-Mix consistently improves downstream tool-use performance across three challenging benchmarks—BFCL, tau2-Bench, and MCP-Universe—under both SFT and RL settings compared to baseline models. The performance gains are particularly pronounced in harder multi-turn, interactive, and out-of-distribution environments, indicating that dedicated mid-training provides a much stronger and more stable substrate for subsequent post-training than raw pre-training alone.

This work shifts the paradigm for developing agentic language models by showing that general tool use is not merely an artifact of post-training instruction tuning, but benefits significantly from foundational exposure to tool dynamics, schema grounding, and workflow recovery during the mid-training phase.