Sign InOpen Brain
arXivPaperNeeds Review

MidTool: Mid-training Data Synthesis for Agentic Tool Use

MidTool trains general tool use before post-training, using API, MCP, document, web, PDF, and code data. Qwen3 4B and 8B variants improved across three downstream tool-use benchmarks.

arXiv · Aug 20, 2026
Open Source Open MarkdownOpen JSON
Source Summary

MidTool builds an open mid-training corpus from web, PDF, and code data plus real-world APIs, MCP skills, and document-grounded workflows. It targets tool selection, argument grounding, workflow composition, and incomplete information.

Practical Implication

The authors mid-trained **Qwen3-4B-Base** and **Qwen3-8B-Base**, then applied both supervised fine-tuning and reinforcement learning. Builders training agent models should consider tool competence a dedicated training stage rather than only a post-training behavior.

Agent-Ready Context
MidTool builds an open mid-training corpus from web, PDF, and code data plus real-world APIs, MCP skills, and document-grounded workflows. It targets tool selection, argument grounding, workflow composition, and incomplete information.

The authors mid-trained **Qwen3-4B-Base** and **Qwen3-8B-Base**, then applied both supervised fine-tuning and reinforcement learning. Builders training agent models should consider tool competence a dedicated training stage rather than only a post-training behavior.

MidTool-Mix improved results under SFT and RL on **BFCL, tau2-Bench, and MCP Universe**, but the supplied material gives no effect sizes. The evidence is also limited to two sizes from one model family.
Connected Context · Feed7 Judgment

MidTool shifts tool competence from an SDK-only or post-training concern into the model’s mid-training curriculum. It supplies open data spanning selection, grounded arguments, workflow composition, and missing information, with improvements under both SFT and RL. The claim remains narrow because effect sizes are absent and evidence covers only Qwen3 at 4B and 8B; it does not establish production reliability.

Muse Spark 1.1 is now available on AI GatewayMuse Spark exposes MCP and parallel tool calling as runtime model features; MidTool supplies a training-stage approach intended to make the underlying selection and composition behavior more competent.MCP Tasks (async): Why Aren't Any Agents Supporting Them? — Cornelia Davis, TemporalMidTool trains workflow composition and incomplete-information handling, while MCP Tasks defines the persistence, retry, cancellation, and recovery contracts needed when those workflows run beyond a connection; model competence does not replace runtime durability.ChromeDevTools/chrome-devtools-mcpChrome DevTools MCP is a concrete, evidence-rich tool surface on which grounded argument choice matters, while its privacy boundary shows that better tool-use training still requires explicit operational controls.Expanding Managed Agents in Gemini API: background tasks, remote MCP and moreGemini Managed Agents provide asynchronous execution, remote MCP, client functions, and credential rotation at runtime; MidTool complements that infrastructure by targeting the model behavior needed to select and compose such tools.
Context Map
modelresearchcoding#tool-use#mcp#agent-sdks
Uncertainty
MidTool-Mix improved results under SFT and RL on **BFCL, tau2-Bench, and MCP Universe**, but the supplied material gives no effect sizes. The evidence is also limited to two sizes from one model family.