Tool-integrated Reinforcement Learning for Repo Deep Search
Zexiong Ma, Chao Peng, Qunhong Zeng, Pengfei Gao, Yanzhen Zou, Bing Xie
2025-08-06
Summary
This paper talks about ToolTrain, a training system that teaches large language models (LLMs) how to use external tools better to find specific issues in large code repositories.
What's the problem?
The problem is that current LLMs have trouble exploring and searching through big codebases effectively because they don't know how to use search tools well and often get lost or make mistakes.
What's the solution?
ToolTrain solves this by training the language model in two stages: first, it learns from good examples of how to use tools correctly; then, it improves through reinforcement learning by trying different tool-using strategies and getting feedback on which ones work best during the search process.
Why it matters?
This matters because it helps AI models understand and navigate complex code faster and more accurately, making it easier to find and fix software bugs or issues, which is very useful for software development.
Abstract
ToolTrain, a two-stage training framework combining supervised fine-tuning and reinforcement learning, enhances LLMs for issue localization by integrating repository retrieval tools, achieving state-of-the-art performance.