MidTool: Mid-training Data Synthesis for Agentic Tool Use

Executive Summary

The capabilities of Large Language Models (LLMs) are evolving at an astonishing pace, yet their path to true agentic intelligence remains fraught with challenges. One increasingly recognized bottleneck is the assumption that complex skills, especially tool use, can be adequately instilled solely through post-training fine-tuning. This paper, “MidTool: Mid-training Data Synthesis for Agentic Tool Use,” provocatively challenges that notion, arguing that general tool use — a cornerstone for effective AI agents — benefits immensely from dedicated mid-training.

Why does this matter right now? As we push for more autonomous and reliable AI agents capable of interacting with the real world, their ability to wield tools effectively and robustly becomes paramount. From navigating APIs to operating complex software, an agent’s utility is directly proportional to its tool mastery. MidTool presents a principled approach to building foundational tool-use skills before final specialization, potentially unlocking a new era of highly capable and adaptable intelligent systems. This work is not just an incremental improvement; it signals a fundamental shift in how we might conceptualize and implement the training pipelines for the next generation of AI.

Technical Deep Dive

The core innovation of MidTool lies in its novel approach to data synthesis for mid-training, a stage of an LLM’s development increasingly understood as critical for shaping fundamental abilities. Unlike pre-training, which establishes broad language understanding, or post-training (fine-tuning, RLHF) which refines behavior for specific tasks, mid-training focuses on instilling specific, complex capabilities after initial pre-training but before extensive task-specific fine-tuning. For agentic tool use, this is a crucial distinction.

The researchers developed an open corpus construction pipeline, also named MidTool, designed to create a rich and diverse dataset for teaching general tool use. This pipeline doesn’t rely solely on existing datasets; it innovatively combines large-scale web, PDF, and code data with synthesized supervision. This synthesis is the key, drawing from several sophisticated sources:

1. Real-world tool APIs: Generating interactions and expected outputs from actual tools.
2. MCP (Multi-level Compositional Planning) skills: Crafting complex scenarios that require sequential and conditional tool invocation.
3. Document-grounded workflows: Simulating tasks where agents must extract information from documents and apply it through tools.

This diverse data enables MidTool to instill a range of critical agentic capabilities:

• Recognizing tool affordances: Understanding what a tool can do and when it’s appropriate to use it.
• Grounding arguments from context: Extracting necessary parameters for tool calls accurately from conversational or task context.
• Composing tool call workflow: Orchestrating multiple tool uses in a logical sequence to achieve a goal.
• Recovering from incomplete information: Handling situations where not all necessary data is present, prompting for it, or inferring it.

The methodology then involved mid-training base models, specifically Qwen3-4B-Base and Qwen3-8B-Base, on this custom MidTool-Mix corpus. Following this, the models underwent standard post-training stages, including both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). The results were compelling: models mid-trained with MidTool-Mix consistently outperformed baselines across various benchmarks, including BFCL, tau2-Bench, and MCP Universe, under both SFT and RL conditions. This robust empirical evidence suggests that general tool use is not a skill best acquired through sheer exposure during post-training but benefits profoundly from dedicated, structured learning during an earlier, formative stage. This underscores the paper’s central thesis: robust agentic tool use requires targeted mid-training.

Real-World Applications

The implications of MidTool’s findings are profound for the development of practical, reliable AI agents across numerous industries:

• Autonomous Software Engineering Agents: Imagine AI agents that can not only generate code but also intelligently use debugging tools, version control systems (like Git), APIs for cloud services, and testing frameworks to autonomously develop, deploy, and maintain software. MidTool could make these agents significantly more robust in identifying and using the correct tools to solve complex programming tasks or fix bugs.
• Advanced Data Analysis and Science: AI agents equipped with MidTool’s enhanced tool-use capabilities could seamlessly interact with statistical software packages (e.g., Pandas, NumPy), visualization libraries (e.g., Matplotlib, Plotly), and even specialized scientific instruments through their APIs. This would empower scientists and analysts to automate complex workflows, accelerate discovery, and derive deeper insights from vast datasets.
• Intelligent Customer Service and Support: Agents could move beyond simple FAQs to leverage CRM systems, ticketing tools, database queries, and external service APIs to resolve complex customer issues, troubleshoot problems, and provide personalized assistance, all while maintaining coherent and effective tool use.
• Robotics and Industrial Automation: For physical AI agents, robust tool use translates to better manipulation of grippers, sensors, and external machinery. MidTool’s principles could lead to agents that are more adept at planning complex sequences of operations, utilizing diverse sensors, and interacting with factory floor control systems, leading to more efficient and adaptable automation.
• Personal AI Assistants: Far beyond current capabilities, a future personal AI agent could genuinely manage your digital life by proficiently using your calendar, email, productivity apps, web search, and specialized services, composing them intelligently to execute complex multi-step requests.

In each of these scenarios, the ability of an LLM to reliably recognize, ground, compose, and recover from tool interactions is the linchpin of its utility. MidTool offers a clearer path to achieving this reliability.

Future Outlook

Looking ahead 2-3 years, the insights from “MidTool: Mid-training Data Synthesis for Agentic Tool Use” will likely catalyze a significant paradigm shift in how we approach the development of intelligent systems. We can anticipate several key trends:

Firstly, the concept of dedicated mid-training stages for critical capabilities will become standard practice, moving beyond general tool use to potentially include areas like complex ethical reasoning, long-term memory management, and advanced planning. The current emphasis on “pre-train, then fine-tune” will evolve into a more nuanced “pre-train, mid-train, then fine-tune” pipeline, leading to more specialized and capable foundation models.

Secondly, the synthesis of high-quality, diverse data for these mid-training stages will itself become a burgeoning field within Machine Learning. Techniques for generating synthetic data from real-world tool APIs, skill graphs, and document-grounded workflows will grow in sophistication, likely leveraging LLMs themselves to generate more realistic and challenging training scenarios. This recursive loop of AI generating data to train better AI will accelerate progress in agentic capabilities.

Thirdly, this enhanced foundational tool use will pave the way for a new generation of truly autonomous and general-purpose AI agents. These won’t be agents narrowly specialized for a single task but rather highly adaptable entities capable of learning new tools on the fly, inferring intent from ambiguous commands, and exhibiting greater resilience in novel environments. The brittleness often seen in current AI agents when faced with slightly out-of-distribution tool-use scenarios will diminish significantly.

Finally, the improved understanding of tool affordances and composition will contribute directly to advancements in explainable AI and human-agent collaboration. Agents that transparently understand why they are choosing a particular tool and how they are using it will be easier to debug, audit, and trust, fostering greater adoption in critical applications. The MidTool paper is not just about tool use; it’s about building the fundamental intelligence required for the future of robust and versatile AI agents.

Key Takeaways

• Mid-training is critical: General agentic tool use, like other complex LLM capabilities, benefits significantly from dedicated mid-training rather than being left solely to post-training.
• MidTool pipeline: The research introduces an open corpus construction pipeline that synthesizes training data from web, PDF, code, real-world tool APIs, MCP skills, and document-grounded workflows.
• Enhanced capabilities: MidTool data is designed to teach models to recognize tool affordances, ground arguments, compose tool workflows, and recover from incomplete information.
• Improved performance: Models like Qwen3-4B-Base and Qwen3-8B-Base mid-trained on MidTool-Mix consistently show superior performance on benchmarks (BFCL, tau2-Bench, MCP Universe) compared to baselines.
• Future of AI agents: This work lays a foundational brick for building more robust, reliable, and general-purpose LLM-powered AI agents capable of effectively interacting with a complex digital and physical world.