PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents

2026-07-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

PhoenixRepair is a novel multi-agent framework designed to enhance automated issue resolution by addressing the limitations of existing Large Language Model-based agents in exploring repair strategies. Current methods exhibit insufficient exploration of both potential edit locations and repair attempts at each location. PhoenixRepair tackles this by systematically exploring multiple candidate edit locations and employing iterative reflection and refinement during patch generation, thereby expanding the search space. Its process involves multi-location sampling, optionally augmented with graph-based localization for complex tasks, followed by iterative patch refinement, and a final generation phase guided by historical insights. Experiments on SWE-bench-Verified demonstrate PhoenixRepair's effectiveness, achieving a 7.8% relative improvement over SWE-agent with DeepSeek-V3.1 and a 76.0% Pass@1 resolved rate using MiniMax-M2.5, alongside improved fault localization accuracy.

Key takeaway

For Machine Learning Engineers or Software Engineers developing automated code repair agents, PhoenixRepair demonstrates that current LLM-based methods often fail due to insufficient exploration of repair strategies. Your teams should consider adopting multi-agent frameworks that systematically explore multiple edit locations and incorporate iterative reflection and refinement for patch generation. This approach, which achieved a 7.8% relative improvement over SWE-agent, can significantly enhance the resolved rate and fault localization accuracy of your automated bug-fixing solutions.

Key insights

PhoenixRepair enhances software agent repair by systematically exploring multiple edit locations and iteratively refining patch generation.

Principles

• Systematic exploration of edit locations
• Iterative reflection refines patch generation
• Distilled insights guide final repair attempts

Method

PhoenixRepair employs multi-location sampling, optionally augmented with graph-based localization, followed by iterative reflection and refinement for patch generation, culminating in a final round guided by historical attempts.

In practice

• Integrate multi-agent exploration for code repair
• Implement iterative reflection for patch improvement
• Apply graph-based localization for difficult bugs

Topics

• Software Agents
• Automated Code Repair
• Large Language Models
• Multi-agent Systems
• Fault Localization
• SWE-bench-Verified

Code references

• DeepSoftwareAnalytics/PhoenixRepair

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• Phoenix: Safe GitHub Issue Resolution via Multi-Agent LLMs — Phoenix prioritizes correctness-first GitHub issue resolution using a multi-agent LLM system with layered safety controls.
• PracRepair: LLM-Empowered Automated Program Repair Inspired by Human-Like Debugging Practices — PracRepair enhances LLM-based automated program repair by integrating dynamic execution and validation feedback, mimicking human debugging.
• Multi-Perspective Agentic Program Repair via Code Property Graphs and Temporal Execution Graphs — Multi-perspective agentic reasoning with structured code and execution graphs significantly improves automated program repair.
• Probe-and-Refine Tuning of Repository Guidance for Coding Agents — Iterative, synthetic bug-fix probing significantly refines LLM coding agent guidance, boosting resolve rates by improving coverage.
• Most Influential ArXiv (Software Engineering) Papers (2026-04 Version) — LLMs and AI agents are rapidly transforming software engineering, driving innovation in code generation, testing, and repair.
• Structured Feedback Improves Repair in an LLM Agent Loop — Structured feedback, particularly with admissible alternatives, substantially improves LLM agent repair success rates.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.