| name | tool-integrated-reasoning-recipe |
| description | Full-pipeline recipe for injecting tool-use into thinking models without sacrificing text-only reasoning. Two-stage SFT+RL approach. Based on arXiv 2605.06326. |
| category | llm-reasoning |
Tool-Integrated Reasoning (TIR) Recipe
Overview
A comprehensive recipe for injecting natural tool-use behavior into strong thinking models without sacrificing text-only reasoning. Achieved 96.7% and 99.2% on AIME 2025 for Qwen3 4B/30B.
Core Pipeline
Stage 1: TIR SFT
- Learnable teacher trajectories: Prioritize problems naturally suited for tool-augmented solutions
- Control tool-use proportion: Balance tool-use vs text-only to mitigate catastrophic forgetting
- Optimize for pass@k and response length: Not training loss — preserves RL headroom
Stage 2: RL with Verifiable Rewards (RLVR)
- Suitable SFT initialization: Start from well-tuned SFT
- Explicit mode collapse safeguards: Prevent over-reliance on tools
- Simple RLVR setup: Effective when built on good SFT foundation
Key Findings
- Tool-enabled evaluation can degrade reasoning even with minimal tool calls
- TIR SFT effectiveness hinges on teacher trajectory learnability
- Tool-use proportion must be controlled to avoid forgetting
- Pass@k and response length optimization > training loss optimization
Implementation Steps
- Curate problems suited for tool augmentation
- Generate teacher trajectories with natural tool-use
- Mix tool-use and text-only trajectories (tune ratio)
- SFT optimizing for pass@k and response length
- RLVR with mode collapse safeguards
- Evaluate on both tool-required and text-only benchmarks
Applicable Use Cases
- Adding tool use to reasoning models
- Preventing catastrophic forgetting
- Multi-stage training pipelines
- Achieving state-of-the-art math reasoning
Triggers / Keywords
tool-integrated reasoning, TIR, SFT+RL pipeline, thinking models, AIME, mode collapse, teacher trajectories
