PhoenixRepair Boosts Software Agent Bug Fixing with Enhanced Strategy Exploration

While Large Language Models (LLMs) have advanced automated software issue resolution, existing agent-based methods often fall short due to insufficient exploration of repair strategies. This limitation manifests in both the number of potential edit locations considered and the depth of repair attempts at each location. Researchers have introduced PhoenixRepair, a multi-agent framework designed to overcome these challenges. It systematically explores multiple candidate edit locations, optionally using graph-based localization for difficult tasks. The framework then employs iterative reflection and refinement to generate better patches, culminating in a final generation guided by distilled insights from all previous attempts. Experiments on SWE-bench-Verified demonstrate PhoenixRepair's effectiveness, showing a 7.8% relative improvement over SWE-agent and achieving the highest resolved rate of 76.0% Pass@1 with MiniMax-M2.5. It also boasts higher fault localization accuracy, indicating a more robust and efficient approach to automated software repair.