MidTool is an open corpus construction pipeline for agentic tool-use mid-training that combines large-scale web, PDF, and code data with synthesized supervision from real-world tool APIs, MCP skills, and document-grounded workflows. It is designed to teach models to recognize tool affordances, ground arguments from context, compose tool call workflows, and recover from incomplete information. Mid-training Qwen3-4B-Base and Qwen3-8B-Base on MidTool-Mix, followed by post-training with SFT and RL, consistently improves downstream performance on BFCL, tau2-Bench, and MCP Universe compared with baselines.
MidTool demonstrates that mid-training with synthesized tool-use supervision can strengthen general tool-use capabilities in LLMs, extending prior mid-training successes in math, science, and software engineering. The pipeline integrates real-world tool APIs and MCP skills, suggesting a scalable approach to teaching tool affordance recognition and workflow composition. Observable next signals include open-sourcing of the MidTool corpus and pipeline, and replication of results on larger models or additional tool-use benchmarks.
This work highlights a growing focus on mid-training as a cost-effective stage for injecting specialized capabilities into LLMs. By targeting general tool use, MidTool addresses a key requirement for agentic AI systems in enterprise and developer contexts. The use of MCP skills and real-world APIs indicates alignment with emerging standards for tool interoperability, potentially accelerating adoption of agentic workflows.
MidTool offers a pathway to improve LLM tool-use performance without full pre-training, potentially reducing costs for developers building agentic applications. Improved performance on benchmarks like BFCL and tau2-Bench suggests practical gains in real-world tool-calling accuracy and workflow execution. The open corpus construction pipeline could lower barriers for enterprises seeking to customize models for proprietary tool ecosystems.
If the MidTool corpus and pipeline are released openly, they could become a foundation for further research and commercial fine-tuning of tool-using models. Subsequent work may explore scaling to larger models, multi-turn tool interactions, and integration with reinforcement learning from environment feedback. The approach may also influence mid-training strategies for other agentic capabilities such as planning and memory.