MidTool: Mid-training Data Synthesis for Agentic Tool Use

Forum topic · 小凯 · 2026-08-22

Summary

MidTool is an open corpus construction pipeline for mid-training large language models on general agentic tool use, introduced in arXiv paper 2608.20314 by Fengqing Jiang, Yite Wang, Boyi Liu, and Zhaoyang Wang. The approach 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, ground arguments in context, compose tool-call workflows, and recover from incomplete information. The authors perform mid-training on Qwen3-4B-Base and Qwen3-8B-Base, followed by supervised fine-tuning (SFT) and reinforcement learning (RL) post-training. Compared with baselines, the resulting MidTool-Mix consistently improves downstream performance on BFCL, tau2-Bench, and MCP Universe benchmarks. The results suggest that general tool use, like other important LLM capabilities, benefits from dedicated mid-training rather than being left entirely to post-training.

Overview

Research area: Machine Learning
Authors: Fengqing Jiang, Yite Wang, Boyi Liu, Zhaoyang Wang
Published: 2026-08-22
arXiv: 2608.20314

Abstract

Mid-training is increasingly recognized as a critical stage for shaping large language model (LLM) capabilities. Recent work has shown that targeted mid-training can enhance reasoning-intensive abilities and 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 corpus construction pipeline for mid-training on agentic tool use. It combines:

Large-scale web, PDF, and code data
Synthetic supervision derived from real-world tool APIs, MCP skills, and documentation-anchored workflows

MidTool is designed to teach models to:

1. Recognize tool functionality 2. Ground arguments in context 3. Compose tool-call workflows 4. Recover from incomplete information

The models are mid-trained 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 dedicated mid-training rather than being left entirely to post-training.

---

*Auto-collected on 2026-08-22.*