Learning Adaptive Parallel Execution for Efficient Code Localization
Summary
FuseSearch is a novel code localization agent developed by Ant Group, Peking University, and Beijing Jiaiao University, designed to optimize the efficiency and quality of identifying relevant code entities for issue resolution. It reformulates parallel code localization as a joint quality-efficiency optimization task, defining tool efficiency as the ratio of unique information gain to invocation count. Utilizing a two-phase SFT and RL training approach, FuseSearch dynamically adjusts search breadth. Evaluated on SWE-bench Verified, FuseSearch-4B achieves 84.7% file-level and 56.4% function-level F₁ scores, demonstrating a 93.6% speedup, 67.7% fewer turns, and 68.9% fewer tokens compared to existing methods. This efficiency-aware training improves quality by eliminating redundant signals.
Key takeaway
For AI Engineers developing automated software development agents, FuseSearch demonstrates that integrating adaptive parallel execution with efficiency-aware training is crucial. You should consider implementing a dual-objective reward function that explicitly penalizes redundant tool invocations, as this not only reduces computational costs by 93.6% but also enhances localization accuracy, enabling faster and more reliable downstream repair tasks.
Key insights
Adaptive parallel execution, guided by tool efficiency, significantly improves code localization speed and accuracy.
Principles
- Tool efficiency quantifies information novelty.
- Joint quality-efficiency optimization is key.
- Adaptive parallelism outperforms fixed breadth.
Method
FuseSearch uses a two-stage SFT and RL training pipeline. SFT bootstraps parallel tool usage with dual-metric filtered trajectories, followed by RL with a multiplicative F₁-efficiency reward to refine adaptive search strategies.
In practice
- Use "grep", "glob", "read_file" for minimalist search.
- Implement dual-metric filtering for SFT data.
- Apply multiplicative F₁-efficiency reward in RL.
Topics
- Code Localization
- Parallel Tool Execution
- Reinforcement Learning
- Software Engineering Agents
- Tool Efficiency
- SWE-bench Verified
Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.