A Survey of LLM-based Automated Program Repair: Taxonomies, Design Paradigms, and Applications

Bach Le; Boyang Yang; Fengling Liu; Haoye Tian; Lingming Zhang; Tegawend\'e F. Bissyand\'e; Yang Liu; Zijian Cai

RECORD

A Survey of LLM-based Automated Program Repair: Taxonomies, Design Paradigms, and Applications

Boyang Yang, Zijian Cai, Fengling Liu, Bach Le, Lingming Zhang, Tegawendé F. Bissyandé, Yang Liu, Haoye Tian

Software Engineering · Submitted 30 Jun 2025

abstract compiled · 2026-09-23 · no language-model review

The authors' main claim This representation exposes control distinctions hidden by utilization labels and supports cross-paradigm analysis of system design and evaluation evidence.

how it is stated · asserted 0.99 read the record →

Comment (0) Read on arXiv ↗

Cite ▾

arXiv ID2506.23749 (version 3)

Pith ID4GDBIFUI, Pith's permanent record of this paper and its review

CitationBach Le, Boyang Yang, Fengling Liu, Haoye Tian, Lingming Zhang, Tegawend\'e F. Bissyand\'e, Yang Liu, Zijian Cai. "A Survey of LLM-based Automated Program Repair: Taxonomies, Design Paradigms, and Applications." arXiv:2506.23749 (2025). https://pith.science/paper/4GDBIFUI

Claim paper

The compiled record

The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Open the compiled record

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 7,563 input tokens.

What the abstract itself states, as probabilities: claims first 0.93 · states limitation 0.30 · human participants 0.04 · controlled comparison 0.06 · randomized assignment 0.01 · states quantitative result 0.90 · compares with prior results 0.35 · mentions code or data availability 0.98.

The abstract, sentence by sentence

arXiv listing abstract v3. Results and the main claim are underlined; show every role.

Large language models (LLMs) are reshaping automated program repair. We present a reproducible hierarchical taxonomy that organizes 66 LLM-based repair systems according to where repair capability and control logic principally reside: task-adapted parameters, prompt and context design, designer-specified workflows, or LLM-directed runtime control. Adaptation, generation pattern, runtime control, and auxiliary evidence are preserved as separate coded dimensions. This representation exposes control distinctions hidden by utilization labels and supports cross-paradigm analysis of system design and evaluation evidence. To the best of our knowledge, it is the first publicly available LLM-based software repair survey to operationalize Fine-Tuning, Prompting, Procedural, and Agentic as one corpus-wide primary classification. Our hierarchy complements prior surveys through one corpus-wide primary decision rule, while the result-level protocol audit bounds which reported results are defensibly comparable. To analyze how repair systems use benchmarks, we record benchmark variants, metrics, base models, and fault-localization assumptions for each system's primary reported result, identifying protocol-aligned comparison windows and benchmark-family fragments. We clarify paradigm trade-offs in task alignment, deployment cost, controllability, and multi-hunk or cross-file repair. We outline open challenges and research directions. Our artifacts and scripted survey pipeline are publicly available at https://github.com/GLEAM-Lab/ProgramRepair.

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	Large language models (LLMs) are reshaping automated program repair.	background 1.00	0.07	0.07	asserted 1.00
s1	We present a reproducible hierarchical taxonomy that organizes 66 LLM-based repair systems according to where repair capability and control logic principally reside: task-adapted parameters, prompt and context design, designer-specified workflows, or LLM-directed runtime control.	contribution 0.98	0.96	0.15	asserted 1.00
s2	Adaptation, generation pattern, runtime control, and auxiliary evidence are preserved as separate coded dimensions.	method 0.96	0.90	0.05	asserted 1.00
s3	This representation exposes control distinctions hidden by utilization labels and supports cross-paradigm analysis of system design and evaluation evidence.	interpretation 0.82	0.85	0.18	asserted 0.99
s4	To the best of our knowledge, it is the first publicly available LLM-based software repair survey to operationalize Fine-Tuning, Prompting, Procedural, and Agentic as one corpus-wide primary classification.	contribution 0.87	0.90	0.12	asserted 0.98
s5	Our hierarchy complements prior surveys through one corpus-wide primary decision rule, while the result-level protocol audit bounds which reported results are defensibly comparable.	contribution 0.93	0.87	0.12	asserted 1.00
s6	To analyze how repair systems use benchmarks, we record benchmark variants, metrics, base models, and fault-localization assumptions for each system's primary reported result, identifying protocol-aligned comparison windows and benchmark-family fragments.	method 0.95	0.95	0.04	asserted 0.98
s7	We clarify paradigm trade-offs in task alignment, deployment cost, controllability, and multi-hunk or cross-file repair.	contribution 0.56	0.85	0.13	asserted 0.97
s8	We outline open challenges and research directions.	future work 0.99	0.79	0.04	asserted 0.67
s9	Our artifacts and scripted survey pipeline are publicly available at https://github.com/GLEAM-Lab/ProgramRepair.	contribution 0.99	0.73	0.02	asserted 1.00

What kind of paper the classifier reports this is: review 1.00 · evidence from literature 0.59.

What was examined, and what was not

sources read

arXiv listing abstract v3

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

The paper

Original abstract

Large language models (LLMs) are reshaping automated program repair. We present a reproducible hierarchical taxonomy that organizes 66 LLM-based repair systems according to where repair capability and control logic principally reside: task-adapted parameters, prompt and context design, designer-specified workflows, or LLM-directed runtime control. Adaptation, generation pattern, runtime control, and auxiliary evidence are preserved as separate coded dimensions. This representation exposes control distinctions hidden by utilization labels and supports cross-paradigm analysis of system design and evaluation evidence. To the best of our knowledge, it is the first publicly available LLM-based software repair survey to operationalize Fine-Tuning, Prompting, Procedural, and Agentic as one corpus-wide primary classification. Our hierarchy complements prior surveys through one corpus-wide primary decision rule, while the result-level protocol audit bounds which reported results are defensibly comparable. To analyze how repair systems use benchmarks, we record benchmark variants, metrics, base models, and fault-localization assumptions for each system's primary reported result, identifying protocol-aligned comparison windows and benchmark-family fragments. We clarify paradigm trade-offs in task alignment, deployment cost, controllability, and multi-hunk or cross-file repair. We outline open challenges and research directions. Our artifacts and scripted survey pipeline are publicly available at https://github.com/GLEAM-Lab/ProgramRepair.

Human review

How Pith checks papers → · Full audit of this paper →

References

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

RECEIPT: Measuring and Repairing the Return Assumption of LLM Agents in Web Service Operations
cs.SE 2026-11 accept novelty 7.0 of 10

When an LLM agent's tool call returns before the program finishes, the harness's input connection silently determines whether the program skips an action, hangs, or proceeds correctly.
On the Reliability of LLM-Based Vulnerability Patching Benchmarks
cs.CR 2026-10 conditional novelty 6.0 of 10

PoC pass rates for LLM security patches swing with prompt and harness choices, and even the newest models fail upstream developer-intent tests on about a third of cases that silence the crash.
PatchyBFT: Automating Diversification of Fault-Tolerant Systems using LLMs
cs.DC 2026-09 conditional novelty 6.0 of 10

LLMs can automatically generate validated, binary-different replacements for up to 65% of the functions in PBFT, HotStuff, and Raft implementations, reducing the cost of protocol diversification.
Code Transformation Rule Synthesis using LLMs: Potential and Limits
cs.SE 2026-09 accept novelty 5.5 of 10

LLMs, especially GPT-5.4, can generate sound and reusable transformation rules in DSLs like Comby, GritQL, and Ast-Grep, but effectiveness drops for complex, multi-statement edits.
Escaping Python Dependency Hell: A Hybrid Replay-and-Repair Pipeline for Python Dependency Resolution
cs.AI 2026-09 conditional novelty 5.0 of 10

On HG2.9K, replaying already-validated dependency configurations produced 1,495 of 1,500 successful repairs, with the LLM fallback contributing only 5.

Show 3 more citing papers

Experimental Settings in LLM-Based Program Repair: A Study of Inputs, Tool Access, Feedback, and Validation
cs.SE 2026-09 accept novelty 5.0 of 10

A framework and schema for specifying the experimental setting of an APR evaluation, demonstrating that benchmark identity alone does not define the repair task.
Revisiting Feedback-Driven LLM Code Repair: A Replication and Exploratory Java Extension
cs.SE 2026-08 conditional novelty 5.0 of 10

A partial replication of FeedbackEval confirms its Python feedback ranking, but an exploratory 100-instance Java extension finds simple feedback as effective as test feedback, suggesting single-language repair conclus...
UML Class Diagram Evaluation and Repair Strategies based on LLMs
cs.SE 2026-08 conditional novelty 5.0 of 10

The paper presents a multidimensional evaluation of LLM-generated UML class diagrams and shows that external knowledge injection, memory reinforcement, and detection-guided repair improve core class coverage to 85% an...

How the literature uses it

23 arXiv papers cite this in 30 sentences

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

Papers citing it · 13 sentences without a confident label

Cites it as background 14 papers

We identified several recent reviews covering related areas, including general LLM-based code generation [ref] , LLMs in software engineering [ref] , code completion [ref] , automated program repair [ref] , parameter-efficient fine-tuning for large code models [ref] , domain-specific and low-resource code generation [ref] , LLMs in programming education [ref] , developer productivity [ref] , and quality and efficiency of LLM-generated code [ref] .

Sustainable Code Generation Using Large Language Models: A Systematic Literature Review · Overview · 0.93
Building on this trend, agentic APR further leverages multiple LLM agents to collaboratively perform localization, patch generation, and validation [ref] .

HELO-APR: Enhancing Low-Resource Program Repair through Cross-Lingual Knowledge Transfer · LLM-Based Program Repair · 0.92
LLM-based APR has evolved through four main paradigms–fine-tuning, prompting, procedural, and agentic [ref] .

BLAgent: Agentic RAG for File-Level Bug Localization · Automated Program Repair · 0.91

What it cites · 85 sentences without a confident label

Uses the method or tool of 5 papers

Each cycle starts with DRMiner [ref] -extracted solution information, letting GPT-4 locate the buggy segment and draft a patch; a tuned CodeT5P [ref] then supplies a reference patch, while an identifier recommender offers renaming hints.

CodeT5+: Open Code Large Language Models for Code Understanding and Generation · Human-in-the-Loop Pipelines · 0.96
Each cycle starts with DRMiner [ref] -extracted solution information, letting GPT-4 locate the buggy segment and draft a patch; a tuned CodeT5P [ref] then supplies a reference patch, while an identifier recommender offers renaming hints.

A Novel Approach for Automated Design Information Mining from Issue Logs · Human-in-the-Loop Pipelines · 0.96
VulMaster fully fine-tunes a CodeT5 [ref] backbone on a security-focused corpus that blends complete vulnerable functions, abstract-syntax-tree structure, and CWE knowledge, then enlists ChatGPT at inference time to supply any missing contextual hints [ref] .

CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and Generation · Full Fine-Tuning · 0.87

Cites as background 13 papers

With the emergence of code-target language models based on transformers [ref] such as DeepSeek-Coder [ref] , CodeT5 [ref] , and CodeLlama [ref] , researchers have pursued full-model supervised fine-tuning on program repair scenarios.

Code Llama: Open Foundation Models for Code · Full Fine-Tuning · 0.92
With the emergence of code-target language models based on transformers [ref] such as DeepSeek-Coder [ref] , CodeT5 [ref] , and CodeLlama [ref] , researchers have pursued full-model supervised fine-tuning on program repair scenarios.

DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence · Full Fine-Tuning · 0.92
With the emergence of code-target language models based on transformers [ref] such as DeepSeek-Coder [ref] , CodeT5 [ref] , and CodeLlama [ref] , researchers have pursued full-model supervised fine-tuning on program repair scenarios.

CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and Generation · Full Fine-Tuning · 0.92

Discussion

Know something about this paper? Leave a comment, or highlight any passage above to comment on it.