Loc2Repair: A Framework for Evaluating the Impact of File-Level Issue Localization in Repo-Level LLM Repair

2026-07-01 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, long

Summary

Loc2Repair is a modular evaluation framework designed to analyze repository-grounded LLM repair pipelines by isolating file-level issue localization as an upstream variable. The framework decouples localization and repair stages under a shared runtime, artifact schema, and evaluation harness, enabling controlled comparisons of different localization models and repair backbones. Evaluating three repair backbones on 500 SWE-bench Verified instances, Loc2Repair demonstrated that explicit localization consistently improves resolved rates. Pooled performance rose from 44.7% for baseline repair to 48.9% with Pred-Qwen4B, 49.1% with Pred-Gemma4E4B, and 52.4% with gold localization. Additionally, localization reduced mean elapsed time, with pooled paired analysis showing decreases of 100.94s, 52.25s, and 154.45s for the respective localization settings, though token usage varied. The study concludes that file-level localization is a significant lever for improving repair effectiveness and latency, yet substantial headroom remains even with oracle-level guidance.

Key takeaway

For MLOps Engineers and Research Scientists designing or evaluating repository-grounded LLM repair agents, you should prioritize integrating a dedicated file-level localization stage. Explicit localization consistently improves resolved rates and often reduces mean elapsed time, as shown by gains from 44.7% to 52.4% with gold guidance. However, recognize that localization is not the sole bottleneck; further decomposition of repair pipelines into specialized stages will yield additional improvements.

Key insights

File-level localization consistently improves LLM-based repository repair effectiveness and latency, but is not the sole bottleneck.

Principles

• Explicit localization enhances LLM repair effectiveness.
• Stage-aware evaluation reveals specific failure modes.
• Localization quality does not monotonically translate to downstream gains.

Method

Loc2Repair decouples localization and repair stages, using a shared runtime and evaluation harness to compare different localizers and repair backbones under matched conditions, isolating localization as an experimental variable.

In practice

• Integrate explicit file localization into LLM repair pipelines.
• Use lightweight localizers for upstream guidance.
• Decompose repair into specialized stages beyond localization.

Topics

• Issue Localization
• Repository-Grounded Repair
• LLM Coding Agents
• SWE-bench
• Evaluation Frameworks
• Software Engineering Agents

Code references

• mohammad-nour-alawad/Loc2Repair

Best for: AI Scientist, Research Scientist, MLOps Engineer

Related on AIssential

• Loc2Repair: A Framework for Evaluating the Impact of File-Level Issue Localization in Repo-Level LLM Repair — File-level localization significantly enhances LLM-based code repair effectiveness and efficiency.
• To Run or Not to Run: Analyzing the Cost-Effectiveness of Code Execution in LLM-Based Program Repair — Code execution in LLM-based program repair often provides minimal benefit despite significant costs, suggesting indiscriminate application.
• Exploration Structure in LLM Agents for Multi-File Change Localization — Linear LLM agent exploration is structurally mismatched for multi-subsystem software changes, benefiting from domain-scoped parallel approaches.
• LLM Agents Can See Code Repositories — Integrating visual structural context with text significantly boosts coding agent efficiency and accuracy in repository tasks.
• Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models — Curvature-guided module localization and low-rank repair effectively detoxify backdoored LLMs post hoc, preserving benign behavior.
• SWE-Explore: Benchmarking How Coding Agents Explore Repositories — SWE-Explore benchmarks coding agents' line-level repository exploration, isolating it from end-to-end repair outcomes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.