MidTool Enhances LLM Agent Tool Use Through Data Synthesis

Mid-training is gaining recognition as a vital phase in shaping the abilities of large language models, particularly for reasoning-intensive tasks and software engineering. This work extends that concept to the less explored area of general tool use for LLM agents. The researchers present MidTool, an open corpus construction pipeline designed to synthesize data specifically for this purpose. MidTool combines extensive web, PDF, and code data with synthesized supervision derived from real-world tool APIs, specific skill sets (MCP skills), and document-grounded workflows. This comprehensive dataset is engineered to teach models how to identify tool affordances, extract arguments from context, compose tool-calling workflows, and recover from incomplete information. The study involved mid-training Qwen3-4B-Base and Qwen3-8B-Base models on MidTool-Mix, followed by post-training using both supervised fine-tuning (SFT) and reinforcement learning. The results consistently showed that MidTool-Mix significantly improved downstream performance across multiple benchmarks, including BFCL, tau2-Bench, and MCP Universe. This strongly suggests that general tool use, much like other critical LLM capabilities, benefits substantially from dedicated mid-training rather than being solely addressed in post-training phases.