Learning Adaptive Parallel Execution for Efficient Code Localization

Researchers introduce FuseSearch, an AI system that optimizes parallel code localization by reducing redundant tool invocations from 34.9% to near-zero through adaptive execution strategies. The approach combines supervised fine-tuning and reinforcement learning to dynamically adjust search breadth, achieving state-of-the-art performance on SWE-bench while using 68.9% fewer tokens and delivering 93.6% speedup.

FuseSearch addresses a critical inefficiency in automated software development pipelines where parallel execution of multiple tools often backfires due to redundant operations. The research reframes code localization from a pure performance problem into a joint quality-efficiency optimization challenge, introducing the metric of tool efficiency—the ratio of unique information gained to total invocations. This conceptual shift enables the system to learn which tools to invoke, when, and in what configuration based on task-specific context rather than applying fixed strategies.

The two-phase training methodology leverages supervised fine-tuning to establish baseline competence before reinforcement learning optimizes for efficiency. FuseSearch dynamically modulates search breadth, transitioning from broad exploration phases when uncertainty is high to narrow refinement stages as the system converges on solutions. This adaptive approach contrasts sharply with static parallel execution methods that waste computational resources on redundant checks.

The empirical results demonstrate that efficiency-aware training naturally improves solution quality by eliminating noisy signals that typically degrade agent performance. Achieving 84.7% file-level accuracy with dramatically reduced token consumption and turn counts indicates the system learns to prioritize high-signal operations. For software development teams and organizations deploying automated debugging tools, this research suggests that resource efficiency and solution quality are not opposing objectives but complementary goals when properly optimized.

Looking forward, the techniques may extend to other domains requiring parallel exploration under resource constraints, including scientific discovery and multi-agent systems. The work validates that training agents with explicit efficiency objectives produces both faster execution and more robust decision-making.

• →FuseSearch reduces redundant tool invocations in parallel code localization from 34.9% to near-zero through adaptive execution learning
• →The system achieves state-of-the-art performance (84.7% file-level F1) while using 68.9% fewer tokens and 67.7% fewer turns
• →Efficiency-aware training improves solution quality by eliminating noisy redundant signals rather than compromising performance for speed
• →Dynamic search breadth adjustment based on task context outperforms fixed-parallelism approaches in both speed and accuracy
• →The methodology combines SFT and RL to train agents that optimize the ratio of unique information gain to tool invocations