Learning Adaptive Parallel Execution for Efficient Code Localization

2026-01-27

AI Summary

This work addresses the inefficiency in code localization caused by redundant tool invocations—currently exhibiting a 34.9% redundancy rate among agents—by proposing FuseSearch, a novel framework that introduces the concept of tool efficiency for the first time. FuseSearch formulates parallel search as a joint optimization task balancing both quality and efficiency. It defines an efficiency metric based on the ratio of information gain to the number of tool calls and employs a two-stage training strategy combining supervised fine-tuning and reinforcement learning, enabling agents to dynamically adjust their search width according to contextual cues rather than adhering to fixed-width constraints. Evaluated on SWE-bench Verified, FuseSearch-4B achieves 84.7% file-level and 56.4% function-level F₁ scores while reducing interaction rounds by 67.7%, token consumption by 68.9%, and attaining a 93.6% speedup.

Technology Category

• Search and Optimization: Learning to Search
• Constraint Satisfaction and Optimization: Search
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Code localization constitutes a key bottleneck in automated software development pipelines. While concurrent tool execution can enhance discovery speed, current agents demonstrate a 34.9\% redundant invocation rate, which negates parallelism benefits. We propose \textbf{FuseSearch}, reformulating parallel code localization as a \textbf{joint quality-efficiency optimization} task. Through defining \textbf{tool efficiency} -- the ratio of unique information gain to invocation count -- we utilize a two-phase SFT and RL training approach for learning adaptive parallel strategies. Different from fixed-breadth approaches, FuseSearch dynamically modulates search breadth according to task context, evolving from exploration phases to refinement stages. Evaluated on SWE-bench Verified, FuseSearch-4B achieves SOTA-level performance (84.7\% file-level and 56.4\% function-level $F_1$ scores) with 93.6\% speedup, utilizing 67.7\% fewer turns and 68.9\% fewer tokens. Results indicate that efficiency-aware training naturally improves quality through eliminating noisy redundant signals, enabling high-performance cost-effective localization agents.

Problem

• code localization
• parallel execution
• redundant invocation
• efficiency
• automated software development

Innovation

• adaptive parallel execution
• tool efficiency
• joint quality-efficiency optimization
• code localization
• reinforcement learning