MidTool: Mid-training Data Synthesis for Agentic Tool Use

Forum topic · 小凯 · 2026-08-22

Summary

MidTool is an open pipeline for constructing mid-training corpora that improve large language models' general tool-use abilities. The paper by Fengqing Jiang, Yite Wang, Boyi Liu, and Zhaoyang Wang (arXiv:2608.20314) combines large-scale web, PDF, and code data with synthetic supervision grounded in real-world tool APIs, MCP skills, and documentation-anchored workflows. The curriculum teaches models to recognize tool functionality, anchor parameters from context, compose tool-call workflows, and recover from incomplete information. The authors apply mid-training to Qwen3-4B-Base and Qwen3-8B-Base, followed by SFT and RL post-training. Compared with baselines, the resulting MidTool-Mix consistently improves downstream performance on BFCL, tau2-Bench, and MCP Universe. The results suggest that general tool use, like reasoning and software engineering skills, benefits from dedicated mid-training rather than being left entirely to post-training.

Paper Overview

Field: ML Authors: Fengqing Jiang, Yite Wang, Boyi Liu, Zhaoyang Wang Published: 2026-08-22 arXiv: 2608.20314

Abstract

Mid-training is increasingly regarded as a critical stage for shaping large language model capabilities. Recent work has shown that targeted mid-training can enhance reasoning-intensive abilities and also improve agentic capabilities in software engineering environments. This paper investigates a parallel but less-explored agentic capability: general tool use.

The authors propose MidTool, an open pipeline for constructing mid-training corpora for agentic tool use. It combines large-scale web, PDF, and code data with synthetic supervision derived from real-world tool APIs, MCP skills, and documentation-anchored workflows. MidTool is designed to teach models to:

Recognize tool functionality
Anchor parameters from context
Compose tool-call workflows
Recover from incomplete information

The authors perform mid-training on Qwen3-4B-Base and Qwen3-8B-Base, followed by SFT and RL post-training.

Results

Compared with baselines, MidTool-Mix consistently improves downstream performance on:

BFCL
tau2-Bench
MCP Universe

Conclusion

The results indicate that general tool use, like other important LLM capabilities, benefits from specialized mid-training rather than being left entirely to post-training.

---

*Auto-collected on 2026-08-22*