Summary

This paper introduces MidTool, the first open pipeline and mid-training dataset specifically designed for general tool use in large language models (LLMs). The authors construct MidTool-Mix, a 20.3B-token corpus that combines web, PDF, code, and structured tool artifacts with two complementary synthesis branches: context-grounded trajectory augmentation and native agentic trajectory synthesis. Through extensive experiments on 4B and 8B base models, they demonstrate that dedicated tool-use mid-training consistently improves downstream performance across three benchmarks (BFCL, τ²-Bench, and MCP-Universe) under both supervised fine-tuning (SFT) and reinforcement learning (RL) settings.

---

1. Introduction and Theoretical Foundation

Background

Tool use is emerging as a defining capability of LLM agents. Effective tool-use systems must:

• Decide when external tools are needed
• Ground calls in tool schemas
• Extract arguments from long and noisy contexts
• Compose multiple tools into workflows
• Recover when information is incomplete

Key Insight

The authors observe that most progress in tool use has come from post-training approaches (SFT and RL on curated traces). However, these approaches place a heavy burden on post-training, as the model must simultaneously acquire multiple atomic capabilities from comparatively narrow supervision. The knowledge underlying successful tool use is distributed across developer documentation, manuals, PDFs, code repositories, API specifications, and structured tool definitions—most of which never appear as clean agent demonstrations.

Research Question

Can general tool-use ability be shaped earlier through dedicated mid-training, rather than being left almost entirely to post-training?

Mid-training is positioned as a distinct stage in multi-stage pre-training, bridging the gap between general pre-training and post-training. Prior work has shown mid-training benefits for reasoning-intensive capabilities (math, science), deep research, and software engineering—but general tool use remained underexplored.

Tool-Use vs. Other Domains

Tool-use data requires covering a broader and more heterogeneous capability surface than math or SWE tasks, including:

• Natural-language documentation
• Executable code patterns
• Structured schemas
• Multi-tool workflows
• Failure cases caused by missing information

---

2. Methodology: The MidTool Pipeline

The pipeline consists of three stages, as illustrated in Figure 2 of the paper.

2.1 Stage 1: Data Source Collection

Four complementary source families are collected:

Key filtering for code data: Known benchmark and evaluation repositories are explicitly excluded through an actively maintained blacklist to prevent leakage.

2.2 Stage 2: Data Preprocessing

Code: Multi-phase filtering adapted from StarCoder, including:

• Extension-based filtering (exclude binaries, model weights, logs)
• Content-based heuristics (line count, line length, alpha ratio)
• Jupyter Notebook conversion to Python
• Deduplication via SHA-256 hashing + MinHash LSH
• Documentation-only directory filtering (docs, examples, tutorials, guides, samples, cookbook)

Web and PDF: Four-phase pipeline:

1. High-recall keyword/URL prescreening
2. Lightweight fastText classifier trained on LLM-labeled seed data
3. Document-level quality filtering
4. MinHash LSH deduplication

2.3 Stage 3: Agentic Trajectory Synthesizing

The synthesis stage addresses two complementary deficits:

Branch A: Context-Grounded Trajectory Augmentation

Target: Grounding deficit—models fail to infer tool boundaries, required arguments, and workflow structure from messy real-world artifacts.

Process:

1. Lightweight keyword-based prefilter
2. Qwen3-235B-A22B-Instruct-2507 annotation with quality score + structured affordance profile
3. Rule-based planner converts extracted affordances into synthesizing plans
4. Budget allocation tied to document quality

Supervision types:

• Regular QA: tool selection, schema-grounded parameter extraction, format-constrained calls, workflow recognition, multiple/parallel use
• Trajectory samples: sequential execution, parameter clarification, tool switching, long-context reasoning

Branch B: Native Agentic Trajectory Synthesis

Target: Execution deficit—models struggle to plan across turns, request missing information, sequence calls, and recover from incomplete interactions.

Process:

1. Build tool inventory by grouping related endpoints/skills
2. GPT-5 assigns quality scores + feasibility profiles
3. Schema refinement for underspecified arguments
4. Deterministic quality-adaptive budget controller (prefers multi-turn trajectories)
5. Category-specific generators (GPT-5, GPT-5.1, GPT-5.2)
6. Strict validation: turn ordering, schema grounding, required arguments, tool-response consistency
7. Mix in rollout trajectories from Agentic World Model (AWM) synthesized environments
8. Incorporate filtered traces from Nemotron Agentic dataset

2.4 MidTool-Mix Composition

Table 2 | Data mixture statistics (tokens in billions):

Slash-separated values denote source corpus / context-grounded augmentation.

All trajectories are normalized into a plain chat-style template without special control tokens (following Olmo et al., 2025).

t-SNE analysis shows MidTool-Mix occupies distinct regions not covered by FineWeb or Dolmino, shifting the distribution toward documentation-heavy, workflow-oriented, and agentic tool-use content. No evidence of benchmark leakage was found.

---

3. Experimental Setup

3.1 Models and Training Recipes

Base models: Qwen3-4B-Base and Qwen3-8B-Base

Four training recipes per scale:

1. Base + SFT (on 100K TOUCAN tool-use subset)
2. Base + SFT + RL
3. Base + MidTool-Mix + SFT
4. Base + MidTool-Mix + SFT + RL

Infrastructure:

• Mid-training and SFT: ArcticTraining on 32 H200 GPUs
• RL: AWM setup with 526 synthetic tool-use environments on 8 B200 GPUs

Benchmarks:

• BFCLv3: Function-calling quality (single-turn and multi-turn)
• τ²-Bench: Interactive task completion (airline, retail, telecom)
• MCP-Universe: Execution over real MCP servers (browser automation, finance, location, web search)

---

4. Key Results

4.1 Main Findings

BFCLv3 Results (Table 3)

Key observation: At 4B scale, multi-turn average rises by more than 10 points over SFT-only, indicating mid-training contributes capabilities not reliably induced by post-training alone.

τ²-Bench Results (Table 4)

MidTool-Mix nearly doubles overall Pass@1 at 4B scale, with strongest gains on airline and retail.

MCP-Universe Results (Table 5)

4.2 Capability Boundary Discovery

Critical finding: Despite substantial improvements on browser automation, financial, and location domains, the web search subset remains at 0.00% across all configurations.

This reveals a meaningful capability boundary:

• General tool-use supervision teaches reusable priors for schema grounding, tool selection, and interaction with unfamiliar APIs
• Highly exploratory domains (deep-search-style tasks) require dedicated trajectory data and training objectives

4.3 Ablation Study (Table 6)

Under fixed 4B SFT recipe, varying only mid-training corpus:

Key insights:

1. Asymmetric contributions: Native agentic trajectories give larger BFCL gains (+7.9 vs. +4.9 overall); context-grounded augmentation is stronger on τ²-Bench and MCP-Universe (transfer)
2. Complementarity: Both single-branch variants fall below no mid-training on MCP-Universe, but the complete mixture improves on all eight metrics
3. Not solely proprietary-driven: The native trajectory branch uses an open-weight teacher; filtered raw data alone still provides positive signal

---

5. Implications

Theoretical Implications

• Mid-training provides a stronger and more stable substrate for subsequent post-training
• General tool-use mid-training and post-training are complementary axes that should advance together
• Capability structure of agentic behavior is non-uniform: broad tool-use priors vs. specialized exploratory behaviors are distinct

Practical Implications

• Dedicated tool-use mid-training can nearly double interactive task completion performance at smaller model scales
• Open-source pipeline enables reproducible research on agentic mid-training
• Provides concrete guidance for future work: search-heavy domains require specialized mid-training data beyond general tool-use mixtures

---

6. Conclusion and Future Directions

Main Takeaways

1. MidTool is the first open pipeline and mid-training dataset for general tool use
2. Dedicated tool-use mid-training consistently improves downstream performance across benchmarks and training recipes
3. Grounding-oriented and execution-oriented supervision are complementary
4. General tool-use mid-training reveals a capability boundary: broad transfer improves, but deep-search-style behaviors remain distinct

Future Work

• Scaling native trajectory collection
• Broadening tool ecosystem coverage
• Constructing specialized mid-training mixtures for domain-specific agentic behaviors (especially search-heavy domains)

---

Key Contributions

1. First open pipeline and mid-training dataset for general tool use (MidTool + MidTool-Mix, 20.3B tokens)
2. Demonstrated consistent improvements across 4B/8B models, three benchmarks, and both SFT and RL settings
3. Revealed a meaningful capability boundary between general tool use and specialized exploratory behaviors, providing guidance for future agentic mid-training research