Learning Adaptive Parallel Execution for Efficient Code Localization

TL;DR

FuseSearch tackles the cost-accuracy trade-off in code localization by learning adaptive parallel execution that explicitly optimizes tool efficiency. It introduces a minimalist, read-only toolset and a two-stage training pipeline (SFT followed by RL) to jointly maximize localization quality and efficiency via a dual-objective reward, $R(\tau) = \alpha F_1(\tau) + \gamma (F_1(\tau) \cdot e(\tau))$. On SWE-bench Verified, FuseSearch achieves state-of-the-art file-level $F_1$ of $84.7\%$ and function-level $F_1$ of $56.4\%$, with a $93.6\%$ speedup and substantial reductions in turns and tokens, while maintaining or improving accuracy over baselines. The approach generalizes to LocBench and demonstrably accelerates downstream tasks (e.g., pre-search localization reduces token usage and time for subsequent repairs), illustrating practical impact in production-grade pipelines. Overall, efficiency-aware training yields high-quality, cost-effective localization by learning when and how to parallelize tool usage across tasks and contexts.

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.

Figures (4)

• Figure 1: Parallel execution solves sequential search's information starvation under limited turns. However, 34.9% of enforced parallel tools are redundant, exhibiting redundancy.
• Figure 2: FuseSearch framework overview. (a) Inference: Agent executes three minimalist tools in parallel, with each tool's information gain tracked to compute trajectory efficiency $e$. (b) Training: Dual-metric filtering selects high-quality trajectories for SFT, followed by RL optimization with joint $F_1$-efficiency reward.
• Figure 3: Evolution of average tools per turn across training stages. RL learns adaptive parallelism: high initial exploration transitioning to focused refinement.
• Figure 4: System prompt for FuseSearch. The prompt instructs the model to output localization results in two sections: Locations to Modify (required) and Related Context (optional).