MidTool: Mid-training Data Synthesis for Agentic Tool Use

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Fengqing Jiang, Yite Wang, Boyi Liu, Zhaoyang Wang, Canwen Xu, Zhewei Yao, Radha Poovendran, Yuxiong He
• Venue: arXiv: cs.AI
• Published: 2026-08-20
• Collected: 2026-08-21
• URL: http://arxiv.org/abs/2608.20314v1
• PDF: https://arxiv.org/pdf/2608.20314v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

In this research, the authors introduce MidTool, a system used for training large language models during the mid-training stage. MidTool focuses on improving the model's abilities in general tool use, which includes recognizing tool affordances, grounding arguments from context, composing tool call workflows, and recovering from incomplete information. The researchers tested MidTool on Qwen3 models and found that it consistently improves downstream performance under both supervised fine-tuning (SFT) and reinforcement learning (RL) on various benchmarks.

What the Document Says

• Mid-training is recognized as critical for shaping large language model capabilities.
• The study focuses on agentic capability: general tool use.
• MidTool is an open corpus construction pipeline for agentic tool-use mid-training that combines various data sources and synthesized supervision from real-world tool APIs, MCP skills, and document-grounded workflows.
• The goal of MidTool is to teach models how to recognize tool affordances, ground arguments from context, compose tool call workflow, and recover from incomplete information.
• The researchers used Qwen3-4B-Base and Qwen3-8B-Base models for mid-training on MidTool-Mix dataset.
• Following mid-training, the models underwent post-training with both supervised fine-tuning and reinforcement learning.
• The study found that MidTool consistently improves downstream performance under SFT and RL on BFCL, tau2-Bench, and MCP Universe.

Key Concepts Mentioned

• mid-training
• large language models
• agentic capability
• general tool use
• MidTool
• Qwen3
• BFCM
• tau2-Bench
• MCP Universe

Claims Made by the Document

• Dedicated mid-training improves general tool use in large language models.
• MidTool is effective in improving downstream performance under SFT and RL.

Evidence Used

• Consistent improvements observed in downstream performance on BFCL, tau2-Bench, and MCP Universe.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Mid-training is increasingly recognized as a critical stage for shaping the capabilities of large language models. Recent work has shown that targeted mid-training can strengthen reasoning-intensive abilities such as math and science, and can also improve agentic capabilities in software-engineering settings. In this work, we study the parallel but less explored agentic capability: general tool use. We present MidTool, 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. MidTool is designed to teach models how to recognize tool affordances, ground arguments from context, compose tool call workflow, and recover from incomplete information. We mid-train Qwen3-4B-Base and Qwen3-8B-Base on MidTool-Mix, and then apply follow-up post-training with both supervised fine-tuning and reinforcement learning. Compared with baselines, MidTool-Mix consistently improves downstream performance under both SFT and RL on BFCL, tau2-Bench, and MCP Universe. These results suggest that general tool use, like other important LLM capabilities, benefits from dedicated mid-training rather than being left entirely to post-training.