Comprehensive survey maps how LLMs are transforming automated code repair

Large language models are fundamentally changing how software engineers approach automated bug fixing. A new comprehensive study from arXiv presents the first major taxonomy to systematically classify these AI-powered repair systems and expose the architectural choices that distinguish them.

Organizing repair into four design paradigms#

The research maps 66 LLM-based repair systems along a structured hierarchy that hinges on where the repair intelligence actually lives. The framework identifies four primary approaches:

• Fine-tuned parameters — systems that adapt pre-trained models through specialized training on repair tasks
• Prompt and context design — systems that steer the model through carefully engineered instructions and surrounding code information
• Procedural workflows — systems with designer-controlled step-by-step repair logic
• Agentic control — systems where the LLM itself drives the repair runtime decisions

This taxonomy reveals control distinctions that simple labels like "prompt-based" or "model-based" had previously obscured, making it easier to compare seemingly similar systems and understand their fundamental trade-offs.

What the survey reveals#

The taxonomy also tracks four cross-cutting dimensions: adaptation strategy, generation pattern, runtime control, and auxiliary evidence (like fault localization hints). By examining how 66 repair systems operate across these facets, the survey identifies substantive differences in deployment cost, task alignment, controllability, and capability at handling complex multi-file or multi-hunk repairs.

The researchers also audit how systems report results, standardizing which findings are truly comparable. They document benchmark variants, metrics, base models, and fault-localization assumptions for each system's primary results—a protocol-level rigor that bounds which performance claims can be fairly matched.

Open questions ahead#

The work outlines emerging research challenges and future directions for the field, while making all artifacts and a reproducible survey pipeline publicly available. This foundational taxonomy is positioned to support cross-paradigm analysis as LLM-based repair continues to evolve.