Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

MidTool: Mid-training Data Synthesis for Agentic Tool Use

Source

arXiv

arxiv.org/abs/2608.20314v1

Read the source

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

Researchers have developed MidTool, a system for teaching large language models how to use tools effectively during mid-training. This is done by combining large-scale web and PDF data with synthetic supervision from real-world tool APIs and document-grounded workflows. The study shows that the performance of Qwen3-4B-Base and Qwen3-8B-Base models improved when trained on MidTool-Mix, a dataset created by this system, compared to baselines in various tasks.

What the Document Says

  • Mid-training is important for shaping the capabilities of large language models.
  • The study focuses on improving agentic capabilities related to general tool use.
  • MidTool is a system that combines large-scale web, PDF, and code data with synthetic supervision from real-world tool APIs, MCP skills, and document-grounded workflows for mid-training.
  • 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 study mid-trains Qwen3-4B-Base and Qwen3-8B-Base on MidTool-Mix.
  • Post-training with both supervised fine-tuning and reinforcement learning is applied after mid-training.
  • The results show that the performance of the models improves when trained on MidTool-Mix compared to baselines in various tasks (BFCL, tau2-Bench, and MCP Universe).

Key Concepts Mentioned

  • mid-training
  • large language models
  • agentic capability
  • general tool use
  • MidTool
  • Qwen3
  • MCP skills
  • document-grounded workflows
  • supervised fine-tuning
  • reinforcement learning

Claims Made by the Document

  • Mid-training is critical for shaping the capabilities of large language models.
  • Targeted mid-training can improve agentic capabilities in software-engineering settings and reasoning-intensive abilities such as math and science.
  • General tool use benefits from dedicated mid-training rather than being left entirely to post-training.

Evidence Used

  • MidTool is presented for improving agentic capabilities related to general tool use.
  • The study shows improved performance of models trained on MidTool-Mix compared to baselines.

Methods / Approach

  • Combining large-scale web, PDF, and code data with synthetic supervision from real-world tool APIs, MCP skills, and document-grounded workflows for mid-training.
  • Post-training with both supervised fine-tuning and reinforcement learning.

Limitations Stated by the Authors

Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

The study does not provide information on the specific baselines used for comparison, or the details of how these baselines were trained.

Open Questions

  • What are the specific baselines used for comparison?
  • How were these baselines trained?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text does not provide detailed information on the baselines used for comparison or their training methods.

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.