Advances and Frontiers of LLM-based Issue Resolution in Software Engineering A Comprehensive Survey
📖 Documentation Website | 📄 Full Paper | 📋 Tables & Resources
🤗 HF Paper: https://huggingface.co/papers/2601.11655 (Upvotes appreciated! ⬆️)
🎙️ Interactive Exploration:
Based on a systematic review of 196 papers and online resources, this survey establishes a holistic theoretical framework for Issue Resolution in software engineering. We examine how Large Language Models (LLMs) are transforming the automation of GitHub issue resolution. Beyond the theoretical analysis, we have curated a comprehensive collection of datasets and model training resources, which are continuously synchronized with our GitHub repository and project documentation website.
🔎 Browse & Export: The full paper database is searchable and exportable at deepsoftwareanalytics.github.io/Awesome-Issue-Resolution/admin/ — filter by category, date, or keyword, and export results as CSV.
2 paper(s) — 2026-03
- BeyondSWE: BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?
- SWE-Adept: SWE-Adept: An LLM-Based Agentic Framework for Deep Codebase Analysis and Structured Issue Resolution
- Survey Update (2026-02): Added 21 new papers covering the latest advances in LLM-based issue resolution!
- Survey Launch (2026-01): Our survey paper is now publicly available on arXiv: https://arxiv.org/abs/2601.11655. It covers 175 papers and resources on LLM-based GitHub issue resolution, with continuously updated datasets and leaderboards!
🔍 Explore This Survey:
- 📊 Data: Evaluation and training datasets, data collection and synthesis methods
- 🛠️ Methods: Training-free (agent/workflow) and training-based (SFT/RL) approaches
- Training-free Methods
- Training-based Methods
- 🔍 Analysis: Insights into both data characteristics and method performance
- 📋 Tables & Resources: Comprehensive statistical tables and resources
- 📄 Full Paper: Read the complete survey paper
- 🤝 Contributing: How to contribute to this project
Total: 196 works across 14 categories
Benchmarks for evaluating issue resolution systems
(2026-02)SWE Context Bench: SWE Context Bench: A Benchmark for Context Learning in Coding(2025-12)SWE-InfraBench: SWE-InfraBench: Evaluating Language Models on Cloud Infrastructure Code(2025-12)SWE-EVO: SWE-EVO: Benchmarking Coding Agents in Long-Horizon Software Evolution Scenarios(2025-11)SWE-Sharp-Bench: SWE-Sharp-Bench: A Reproducible Benchmark for C# Software Engineering Tasks(2025-11)SWE-fficiency: SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads?(2025-11)SWE-Compass: SWE-Compass: Towards Unified Evaluation of Agentic Coding Abilities for Large Language Models(2025-09)SWE-Bench Pro: SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?(2025-07)SWE-Perf: SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?(2025-05)SwingArena: SwingArena: Competitive Programming Arena for Long-context GitHub Issue Solving(2025-05)OmniGIRL: Omnigirl: A multilingual and multimodal benchmark for github issue resolution(2025-05)SWE-bench-Live: SWE-bench Goes Live!(2025-04)Multi-SWE-bench: Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving(2025-04)SWE-PolyBench: SWE-PolyBench: A multi-language benchmark for repository level evaluation of coding agents(2025-04)SWE-bench Multilingual: SWE-smith: Scaling Data for Software Engineering Agents(2025-03)FEA-Bench: FEA-Bench: A Benchmark for Evaluating Repository-Level Code Generation for Feature Implementation(2025-02)SWE-Lancer: SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering?(2024-12)Visual SWE-bench: CodeV: Issue Resolving with Visual Data(2024-10)SWE-bench Multimodal: SWE-bench Multimodal: Do AI Systems Generalize to Visual Software Domains?(2024-08)SWE-bench-java: SWE-bench-java: A GitHub Issue Resolving Benchmark for Java
Datasets for training issue resolution agents
(2026-02)SWE-Universe: SWE-Universe: Scale Real-World Verifiable Environments to Millions(2025-06)Skywork-SWE: Skywork-SWE: Unveiling Data Scaling Laws for Software Engineering in LLMs(2025-05)SWELoc: SweRank: Software Issue Localization with Code Ranking(2025-04)Multi-SWE-RL: Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving(2025-04)SWE-Smith: SWE-smith: Scaling Data for Software Engineering Agents(2025-02)LocAgent: OrcaLoca: An LLM Agent Framework for Software Issue Localization(2025-01)SWE-Fixer: SWE-Fixer: Training Open-Source LLMs for Effective and Efficient GitHub Issue Resolution(2023-10)SWE-bench-extra: SWE-bench: Can Language Models Resolve Real-world Github Issues?
Individual autonomous agents for issue resolution
(2025-12)Confucius Code Agent: Confucius Code Agent: Scalable Agent Scaffolding for Real-World Codebases(2025-10)TOM-SWE: TOM-SWE: User Mental Modeling For Software Engineering Agents(2025-09)Lita: Lita: Light Agent Uncovers the Agentic Coding Capabilities of LLMs(2025-08)Live-SWE-agent: SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents(2025-07)Trae Agent: Trae Agent: An LLM-based Agent for Software Engineering with Test-time Scaling(2025-05)LCLM: Putting It All into Context: Simplifying Agents with LCLMs(2025-02)PatchPilot: PatchPilot: A Cost-Efficient Software Engineering Agent with Early Attempts on Formal Verification(2024-05)SWE-agent: Swe-agent: Agent-computer interfaces enable automated software engineering(2024-03)Devin: SWE-bench technical report(2023-06)Aider
Collaborative multi-agent frameworks
(2025-08)Meta-RAG: Meta-RAG on Large Codebases Using Code Summarization(2025-07)SWE-Debate: SWE-Debate: Competitive Multi-Agent Debate for Software Issue Resolution(2025-06)AgentScope: SWE-Bench - AgentScope(2025-05)Devlo: Achieving SOTA on SWE-bench(2025-05)Refact.ai Agent: AI Coding Agent for Software Development - Refact.ai(2025-03)Lingxi: Lingxi/docs/Lingxi Technical Report 2505.pdf at master · lingxi-agent/Lingxi(2025-02)OrcaLora: OrcaLoca: An LLM Agent Framework for Software Issue Localization(2025-01)CodeCoR: CodeCoR: An LLM-Based Self-Reflective Multi-Agent Framework for Code Generation(2024-09)MarsCode Agent: MarsCode Agent: AI-native Automated Bug Fixing(2024-09)HyperAgent: HyperAgent: Generalist Software Engineering Agents to Solve Coding Tasks at Scale(2024-08)DEI: Diversity Empowers Intelligence: Integrating Expertise of Software Engineering Agents(2024-07)OpenHands: OpenHands: An Open Platform for AI Software Developers as Generalist Agents(2024-06)CodeR: CodeR: Issue Resolving with Multi-Agent and Task Graphs(2024-04)AutoCodeRover: AutoCodeRover: Autonomous Program Improvement(2024-03)MAGIS: MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution
Structured pipeline approaches
(2025-07)SynFix: SynFix: Dependency-Aware Program Repair via RelationGraph Analysis(2025-06)GUIRepair: Seeing is Fixing: Cross-Modal Reasoning with Multimodal LLMs for Visual Software Issue Fixing(2024-12)CodeV: CodeV: Issue Resolving with Visual Data(2024-10)Conversational Pipeline: Exploring the Potential of Conversational Test Suite Based Program Repair on SWE-bench(2024-07)Agentless: Demystifying LLM-Based Software Engineering Agents
Methods leveraging external tools
(2026-02)Closing the Loop: Closing the Loop: Universal Repository Representation with RPG-Encoder(2026-01)SWE-Tester: SWE-Tester: Training Open-Source LLMs for Issue Reproduction in Real-World Repositories(2025-12)GraphLocator: GraphLocator: Graph-guided Causal Reasoning for Issue Localization(2025-11)InfCode: InfCode: Adversarial Iterative Refinement of Tests and Patches for Reliable Software Issue Resolution(2025-10)BugPilot: BugPilot: Complex Bug Generation for Efficient Learning of SWE Skills(2025-10)TestPrune: When Old Meets New: Evaluating the Impact of Regression Tests on SWE Issue Resolution(2025-09)Nemotron-CORTEXA: Nemotron-CORTEXA: Enhancing LLM Agents for Software Engineering Tasks via Improved Localization and Solution Diversity(2025-08)Git Context Controller: Git Context Controller: Manage the Context of LLM-based Agents like Git(2025-07)Prometheus: Prometheus: Unified Knowledge Graphs for Issue Resolution in Multilingual Codebases(2025-06)SACL: SACL: Understanding and Combating Textual Bias in Code Retrieval with Semantic-Augmented Reranking and Localization(2025-06)OpenHands-Versa: Coding Agents with Multimodal Browsing are Generalist Problem Solvers(2025-06)SemAgent: SemAgent: A Semantics Aware Program Repair Agent(2025-06)Repeton: Repeton: Structured Bug Repair with ReAct-Guided Patch-and-Test Cycles(2025-06)cAST: cAST: Enhancing Code Retrieval-Augmented Generation with Structural Chunking via Abstract Syntax Tree(2025-05)InfantAgent-Next: InfantAgent-Next: A Multimodal Generalist Agent for Automated Computer Interaction(2025-05)SWERank: SweRank: Software Issue Localization with Code Ranking(2025-03)DARS: DARS: Dynamic Action Re-Sampling to Enhance Coding Agent Performance by Adaptive Tree Traversal(2025-03)Issue2Test: Issue2Test: Generating Reproducing Test Cases from Issue Reports(2025-03)KGCompass: Enhancing repository-level software repair via repository-aware knowledge graphs(2025-03)CoSIL: Issue Localization via LLM-Driven Iterative Code Graph Searching(2025-02)OrcaLoca: OrcaLoca: An LLM Agent Framework for Software Issue Localization(2025-02)Otter: Otter: Generating Tests from Issues to Validate SWE Patches(2025-02)Quadropic Insiders: Quadropic Insiders : Syntheo Tops Swelite Feb(2024-12)CoRNStack: CoRNStack: High-Quality Contrastive Data for Better Code Retrieval and Reranking(2024-11)AEGIS: AEGIS: An Agent-based Framework for General Bug Reproduction from Issue Descriptions(2024-10)RepoGraph: RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph(2024-09)SuperCoder2.0: SuperCoder2.0: Technical Report on Exploring the feasibility of LLMs as Autonomous Programmer(2024-08)SpecRover: SpecRover: Code Intent Extraction via LLMs(2024-06)Alibaba LingmaAgent: Alibaba LingmaAgent: Improving Automated Issue Resolution via Comprehensive Repository Exploration
Systems with memory mechanisms
(2026-01)MemGovern: MemGovern: Enhancing Code Agents through Learning from Governed Human Experiences(2025-10)RepoMem: Improving Code Localization with Repository Memory(2025-09)AgentDiet: Improving the Efficiency of LLM Agent Systems through Trajectory Reduction(2025-07)Agent KB: Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving(2025-07)SWE-Exp: SWE-Exp: Experience-Driven Software Issue Resolution(2025-06)ExpeRepair: EXPEREPAIR: Dual-Memory Enhanced LLM-based Repository-Level Program Repair(2025-05)DGM: Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents(2024-11)Infant Agent: Infant Agent: A Tool-Integrated, Logic-Driven Agent with Cost-Effective API Usage(2024-11)EvoCoder: LLMs as Continuous Learners: Improving the Reproduction of Defective Code in Software Issues
Models trained via supervised learning
(2026-01)SWE-Lego: SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving(2026-01)SWE-Replay: SWE-Replay: Efficient Test-Time Scaling for Software Engineering Agents(2025-12)SWE-Compressor: Context as a Tool: Context Management for Long-Horizon SWE-Agents(2025-09)Devstral: Devstral: Fine-tuning Language Models for Coding Agent Applications(2025-06)MCTS-Refined CoT: MCTS-Refined CoT: High-Quality Fine-Tuning Data for LLM-Based Repository Issue Resolution(2025-05)Search for training: Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents(2025-05)Co-PatcheR: Co-PatcheR: Collaborative Software Patching with Component(s)-specific Small Reasoning Models(2025-05)CGM: Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks(2025-03)Thinking Longer: Thinking Longer, Not Larger: Enhancing Software Engineering Agents via Scaling Test-Time Compute(2024-12)ReSAT: Repository Structure-Aware Training Makes SLMs Better Issue Resolver(2024-12)Scaling data collection: Scaling Data Collection for Training SWE Agents(2024-12)SWE-Gym: Training Software Engineering Agents and Verifiers with SWE-Gym(2024-11)Lingma SWE-GPT: SWE-GPT: A Process-Centric Language Model for Automated Software Improvement(2024-11)CodeXEmbed: CodeXEmbed: A Generalist Embedding Model Family for Multilingual and Multi-task Code Retrieval
Models trained via reinforcement learning
(2026-02)SWE-Master: SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training(2026-02)SWE-Protégé: SWE-Protégé: Learning to Selectively Collaborate With an Expert Unlocks Small Language Models as Software Engineering Agents(2026-02)SWE-MiniSandbox: SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents(2026-01)MiMo-V2-Flash: MiMo-V2-Flash Technical Report(2025-12)Self-play SWE-RL: Toward Training Superintelligent Software Agents through Self-Play SWE-RL(2025-12)SWE-Playground: Training Versatile Coding Agents in Synthetic Environments(2025-12)SWE-RM: SWE-RM: Execution-free Feedback For Software Engineering Agents(2025-12)One Tool Is Enough: One Tool Is Enough: Reinforcement Learning for Repository-Level LLM Agents(2025-12)Let It Flow: Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem(2025-12)Deepseek V3.2: DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models(2025-11)TSP: Think-Search-Patch: A Retrieval-Augmented Reasoning Framework for Repository-Level Code Repair(2025-10)CWM: CWM: An Open-Weights LLM for Research on Code Generation with World Models(2025-10)FoldGRPO: Scaling Long-Horizon LLM Agent via Context-Folding(2025-10)GRPO-based Method: A Practitioner's Guide to Multi-turn Agentic Reinforcement Learning(2025-10)Supervised RL: Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning(2025-10)KAT-Coder: KAT-Coder Technical Report(2025-09)CoreThink: CoreThink: A Symbolic Reasoning Layer to reason over Long Horizon Tasks with LLMs(2025-09)EntroPO: Building Coding Agents via Entropy-Enhanced Multi-Turn Preference Optimization(2025-09)Kimi-Dev: Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents(2025-09)LongCat-Flash-Think: Introducing LongCat-Flash-Thinking: A Technical Report(2025-08)Tool-integrated RL: Tool-integrated Reinforcement Learning for Repo Deep Search(2025-08)SWE-Swiss: SWE-Swiss: A Multi-Task Fine-Tuning and RL Recipe for High-Performance Issue Resolution(2025-08)SeamlessFlow: SeamlessFlow: A Trainer Agent Isolation RL Framework Achieving Bubble-Free Pipelines via Tag Scheduling(2025-08)DAPO: Training Long-Context, Multi-Turn Software Engineering Agents with Reinforcement Learning(2025-08)GLM-4.6: gpt-oss-120b & gpt-oss-20b model card(2025-07)DeepSWE: DeepSWE: Training a State-of-the-Art Coding Agent from Scratch by Scaling RL(2025-07)Kimi-K2-Instruct: Kimi K2: Open Agentic Intelligence(2025-06)Agent-RLVR: Agent-RLVR: Training Software Engineering Agents via Guidance and Environment Rewards(2025-06)SWE-Dev2: SWE-Dev: Building Software Engineering Agents with Training and Inference Scaling(2025-06)Minimax M2: MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention(2025-05)SWE-Dev1: SWE-Dev: Evaluating and Training Autonomous Feature-Driven Software Development(2025-05)Satori-SWE: Satori-SWE: Evolutionary Test-Time Scaling for Sample-Efficient Software Engineering(2025-05)Qwen3-Coder: Qwen3 Technical Report(2025-04)Seed1.5-Thinking: Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning(2025-03)SEAlign: SEAlign: Alignment Training for Software Engineering Agent(2025-02)SWE-RL: SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution(2025-02)SoRFT: SoRFT: Issue Resolving with Subtask-oriented Reinforced Fine-Tuning(2024-10)OSCA: Scaling LLM Inference Efficiently with Optimized Sample Compute Allocation
Methods for scaling at inference time
(2026-01)Agentic Rubrics: Agentic Rubrics as Contextual Verifiers for SWE Agents(2025-10)SIADAFIX: SIADAFIX: issue description response for adaptive program repair(2025-09)SWE-PRM: When Agents go Astray: Course-Correcting SWE Agents with PRMs(2025-01)ReasoningBank: CodeMonkeys: Scaling Test-Time Compute for Software Engineering(2024-10)SWE-Search: SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement
Techniques for collecting training data
(2026-02)DockSmith: DockSmith: Scaling Reliable Coding Environments via an Agentic Docker Builder(2026-01)MEnvAgent: MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering(2025-12)Multi-Docker-Eval: Multi-Docker-Eval: A `Shovel of the Gold Rush' Benchmark on Automatic Environment Building for Software Engineering(2025-08)RepoForge: RepoForge: Training a SOTA Fast-thinking SWE Agent with an End-to-End Data Curation Pipeline Synergizing SFT and RL at Scale(2025-07)SWE-MERA: SWE-MERA: A Dynamic Benchmark for Agenticly Evaluating Large Language Models on Software Engineering Tasks(2025-06)SWE-Factory: SWE-Factory: Your Automated Factory for Issue Resolution Training Data and Evaluation Benchmarks(2025-05)SWE-rebench: SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents(2025-05)RepoLaunch: SWE-bench Goes Live!
Approaches for synthetic data generation
(2026-02)SWE-World: SWE-World: Building Software Engineering Agents in Docker-Free Environments(2025-09)SWE-Mirror: SWE-Mirror: Scaling Issue-Resolving Datasets by Mirroring Issues Across Repositories(2025-06)SWE-Flow: Synthesizing Software Engineering Data in a Test-Driven Manner(2025-04)R2E-Gym: R2E-Gym: Procedural Environment Generation and Hybrid Verifiers for Scaling Open-Weights SWE Agents(2025-04)SWE-Synth: SWE-Synth: Synthesizing Verifiable Bug-Fix Data to Enable Large Language Models in Resolving Real-World Bugs(2025-04)SWE-smith: SWE-smith: Scaling Data for Software Engineering Agents(2025-01)Learn-by-interact: Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments
Analysis of datasets and benchmarks
(2025-12)Data contamination: Does SWE-Bench-Verified Test Agent Ability or Model Memory?(2025-07)Rigorous agentic benchmarks: Establishing Best Practices for Building Rigorous Agentic Benchmarks(2025-07)SPICE: SPICE: An Automated SWE-Bench Labeling Pipeline for Issue Clarity, Test Coverage, and Effort Estimation(2025-06)UTBoost: UTBoost: Rigorous Evaluation of Coding Agents on SWE-Bench(2025-06)Trustworthiness: Is Your Automated Software Engineer Trustworthy?(2025-06)The SWE-Bench Illusion: The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reason(2025-04)Revisiting SWE-Bench: Revisiting SWE-Bench: On the Importance of Data Quality for LLM-Based Code Models(2025-03)Patch Correctness: Are "Solved Issues" in SWE-bench Really Solved Correctly? An Empirical Study(2024-08)SWE-bench Verified: Introducing SWE-bench Verified | OpenAI
Comparative analysis of different methods
(2025-12)SWEnergy: SWEnergy: An Empirical Study on Energy Efficiency in Agentic Issue Resolution Frameworks with SLMs(2025-09)Failures analysis: An Empirical Study on Failures in Automated Issue Solving(2025-07)Security analysis: How Safe Are AI-Generated Patches? A Large-scale Study on Security Risks in LLM and Agentic Automated Program Repair on SWE-bench(2025-06)Dissecting the SWE-Bench Leaderboards: Dissecting the SWE-Bench Leaderboards: Profiling Submitters and Architectures of LLM- and Agent-Based Repair Systems(2025-05)GSO: GSO: Challenging Software Optimization Tasks for Evaluating SWE-Agents(2025-05)Strong-Weak Model Collaboration: An Empirical Study on Strong-Weak Model Collaboration for Repo-level Code Generation(2025-05)Agents in the Wild(2025-04)SeaView: SeaView: Software Engineering Agent Visual Interface for Enhanced Workflow(2025-03)Beyond final code: Beyond Final Code: A Process-Oriented Error Analysis of Software Development Agents in Real-World GitHub Scenarios(2025-02)Overthinking: The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks(2024-10)Evaluating software development agents: Evaluating Software Development Agents: Patch Patterns, Code Quality, and Issue Complexity in Real-World GitHub Scenarios(2024-06)Context Retrieval: On The Importance of Reasoning for Context Retrieval in Repository-Level Code Editing
A comprehensive survey and statistical overview of issue resolution datasets. We categorize these datasets based on programming language, modality support, source repositories, data scale (Amount), and the availability of reproducible execution environments.
| Dataset | Language | Multimodal | Repos | Amount | Environment | Link |
|---|---|---|---|---|---|---|
| Single-PL Datasets | ||||||
| SWE-Fixer | Python | No | 856 | 115,406 | No |   |
| SWE-smith | Python | No | 128 | 50k | Yes | |
| SWE-Lego | Python | No | 3,251 | 32,119 | Yes | |
| SWE-rebench | Python | No | 3,468 | 21,336 | Yes | |
| SWE-bench-train | Python | No | 37 | 19k | No | |
| SWE-Flow | Python | No | 74 | 18,081 | Yes | |
| Skywork-SWE | Python | No | 2,531 | 10,169 | Yes | - |
| R2E-Gym | Python | No | 10 | 8,135 | Yes | |
| RepoForge | Python | No | - | 7.3k | Yes | - |
| SWE-bench-extra | Python | No | 2k | 6.38k | Yes | |
| SWE-Gym | Python | No | 11 | 2,438 | Yes | |
| SWE-bench | Python | No | 12 | 2,294 | Yes | |
| SWE-bench-java | Java | No | 19 | 1,797 | Yes | |
| FEA-bench | Python | No | 83 | 1,401 | Yes | |
| SWE-bench-Live | Python | No | 164 | 1,565 | Yes | |
| Loc-Bench | Python | No | - | 560 | No | |
| SWE-bench Verified | Python | No | - | 500 | Yes | |
| SWE-bench Lite | Python | No | 12 | 300 | Yes | |
| SWE-MERA | Python | No | 200 | 300 | Yes | |
| SWE-Bench-CL | Python | No | 8 | 273 | Yes | |
| SWE-Sharp-Bench | C# | No | 17 | 150 | Yes | |
| SWE-Perf | Python | No | 12 | 140 | Yes | |
| Visual SWE-bench | Python | Yes | 11 | 133 | Yes | |
| SWE-EVO | Python | No | 7 | 48 | Yes | |
| Multi-PL Datasets | ||||||
| SWE-Mirror | Python, Rust, Go | No | 40 | 60k | Yes | - |
| Multi-SWE-bench | Java, JS, TS, Go, Rust, C, C++ | No | 76 | 4,723 | Yes | |
| Swing-Bench | Python, Go, C++, Rust | No | 400 | 2300 | Yes | - |
| SWE-PolyBench | Python, Java, JS, TS | No | 21 | 2,110 | Yes | |
| SWE-Compass | Python, JS, TS, Java, C, C++, Go, Rust, Kotlin, C# | No | - | 2,000 | Yes | |
| SWE-Bench Pro | Python, Go, TS | No | 41 | 1,865 | Yes | |
| SWE-bench++ | Python, Go, TS, JS, Ruby, PHP, Java, Rust, C++, C#, C | No | 3,971 | 1,782 | Yes | |
| SWE-Lancer | JS, TS | No | - | 1,488 | Yes | |
| OmniGIRL | Python, TS, Java, JS | Yes | 15 | 959 | Yes | |
| SWE-bench Multimodal | JS, TS, HTML, CSS | Yes | 17 | 619 | Yes | |
| SWE-fficiency | Python, Cython | No | 9 | 498 | Yes | |
| SWE-Factory | Python, Java, JS, TS | No | 12 | 430 | Yes | |
| SWE-bench-Live-MultiLang & Windows | Python, JS, TS, C, C++, C#, Java, Go, Rust | No | 238 | 418 | Yes | |
| SWE-bench Multilingual | C, C++, Go, Java, JS, TS, Rust, Python, Ruby, PHP | No | 42 | 300 | Yes | |
| SWE-InfraBench | Python, TS | No | - | 100 | Yes | - |
A survey of trajectory datasets used for agent training or analysis. We list the programming language, number of source repositories, and total trajectories for each dataset.
| Dataset | Language | Repos | Amount | Link |
|---|---|---|---|---|
| SWE-Fixer | Python | 856 | 69,752 | |
| SWE-rebench | Python | 1,823 | 67,074 | |
| R2E-Gym | Python | 10 | 3,321 | |
| SWE-Synth | Python | 11 | 3,018 | |
| SWE-Factory | Python | 10 | 2,809 | |
| SWE-Gym | Python | 11 | 491 | |
| SWE-Lego | Python | 3251 | 14.6k | |
Overview of SFT-based methods for issue resolution. This table categorizes models by their base architecture and training scaffold (Sorted by Performance).
| Model Name | Base Model | Size | Arch. | Training Scaffold | Res.(%) | Code | Data | Model |
|---|---|---|---|---|---|---|---|---|
| SWE-rebench-openhands-Qwen3-235B-A22B | Qwen3-235B-A22B | 235B-A22B | MoE | OpenHands | 59.9 | - | |
|
| SWE-Lego-Qwen3-32B | Qwen3-32B | 32B | Dense | OpenHands | 57.6 | |
|
|
| CGM-SWE-PY | Qwen2.5-Coder-72B | 72B | Dense | Graph RAG | 50.4 | |
- | |
| SWE-rebench-openhands-Qwen3-30B-A3B | Qwen3-30B-A3B | 30B-A3B | MoE | OpenHands | 49.7 | - | |
|
| Devstral | Mistral Small 3 | 22B | Dense | OpenHands | 46.8 | - | - | |
| Co-PatcheR | Qwen2.5-Coder-14B | 3$\times$14B | Dense | PatchPilot-mini | 46.0 | |
- | |
| SWE-Swiss-32B | Qwen2.5-32B-Instruct | 32B | Dense | Agentless | 45.0 | |
|
|
| SWE-Lego-Qwen3-8B | Qwen3-8B | 8B | Dense | OpenHands | 44.4 | |
|
|
| Lingma SWE-GPT | Qwen2.5-72B-Instruct | 72B | Dense | SWESynInfer | 30.2 | |
- | - |
| SWE-Gym-Qwen-32B | Qwen2.5-Coder-32B | 32B | Dense | OpenHands, MoatlessTools | 20.6 | |
- | |
| SWE-Gym-Qwen-14B | Qwen2.5-Coder-14B | 14B | Dense | OpenHands, MoatlessTools | 16.4 | |
- | |
| SWE-Gym-Qwen-7B | Qwen2.5-Coder-7B | 7B | Dense | OpenHands, MoatlessTools | 10.6 | |
- | |
A comprehensive overview of specialized models for issue resolution, categorized by parameter size. The table details each model's base architecture, the training scaffold used for rollout, the type of reward signal employed (Outcome vs. Process), and their performance results (Res. %) on issue resolution benchmarks.
| Model Name | Base Model | Size | Arch. | Train. Scaffold | Reward | Res.(%) | Code | Data | Model |
|---|---|---|---|---|---|---|---|---|---|
| 560B Models (MoE) | |||||||||
| LongCat-Flash-Think | LongCatFlash-Base | 560B-A27B | MoE | R2E-Gym | Outcome | 60.4 | |
- | |
| 72B Models | |||||||||
| Kimi-Dev | Qwen 2.5-72B-Base | 72B | Dense | BugFixer + TestWriter | Outcome | 60.4 | |
- | |
| SWE-RL | Llama-3.3-70B-Instruct | 70B | Dense | Agentless-mini | Outcome | 41.0 | |
- | - |
| Multi-turn RL(Nebius) | Qwen2.5-72B-Instruct | 72B | Dense | SWE-agent | Outcome | 39.0 | - | - | - |
| Agent-RLVR-RM-72B | Qwen2.5-Coder-72B | 72B | Dense | Localization + Repair | Outcome | 27.8 | - | - | - |
| Agent-RLVR-72B | Qwen2.5-Coder-72B | 72B | Dense | Localization + Repair | Outcome | 22.4 | - | - | - |
| 32B Models | |||||||||
| OpenHands Critic | Qwen2.5-Coder-32B | 32B | Dense | SWE-Gym | - | 66.4 | |
- | |
| KAT-Dev-32B | Qwen3-32B | 32B | Dense | - | - | 62.4 | - | - | |
| SWE-Swiss-32B | Qwen2.5-32B-Instruct | 32B | Dense | - | Outcome | 60.2 | |
|
|
| FoldAgent | Seed-OSS-36B-Instruct | 36B | Dense | FoldAgent | Process | 58.0 | |
- | - |
| SeamlessFlow-32B | Qwen3-32B | 32B | Dense | SWE-agent | Outcome | 45.8 | |
- | - |
| DeepSWE | Qwen3-32B | 32B | Dense | R2E-Gym | Outcome | 42.2 | |
|
|
| SA-SWE-32B | - | 32B | Dense | SkyRL-Agent | - | 39.4 | - | - | - |
| OpenHands LM v0.1 | Qwen2.5-Coder-32B | 32B | Dense | SWE-Gym | - | 37.2 | |
- | |
| SWE-Dev-32B | Qwen2.5-Coder-32B | 32B | Dense | OpenHands | Outcome | 36.6 | |
- | |
| Satori-SWE | Qwen2.5-Coder-32B | 32B | Dense | Retriever + Code editor | Outcome | 35.8 | |
|
|
| SoRFT-32B | Qwen2.5-Coder-32B | 32B | Dense | Agentless | Outcome | 30.8 | - | - | - |
| Agent-RLVR-32B | Qwen2.5-Coder-32B | 32B | Dense | Localization + Repair | Outcome | 21.6 | - | - | - |
| 14B Models | |||||||||
| Agent-RLVR-14B | Qwen2.5-Coder-14B | 14B | Dense | Localization + Repair | Outcome | 18.0 | - | - | - |
| SEAlign-14B | Qwen2.5-Coder-14B | 14B | Dense | OpenHands | Process | 17.7 | - | - | - |
| 7-8B Models | |||||||||
| SeamlessFlow-8B | Qwen3-8B | 8B | Dense | SWE-agent | Outcome | 27.4 | |
- | - |
| SWE-Dev-7B | Qwen2.5-Coder-7B | 7B | Dense | OpenHands | Outcome | 23.4 | |
- | |
| SoRFT-7B | Qwen2.5-Coder-7B | 7B | Dense | Agentless | Outcome | 21.4 | - | - | - |
| SWE-Dev-8B | Llama-3.1-8B | 8B | Dense | OpenHands | Outcome | 18.0 | |
- | |
| SEAlign-7B | Qwen2.5-Coder-7B | 7B | Dense | OpenHands | Process | 15.0 | - | - | - |
| SWE-Dev-9B | GLM-4-9B | 9B | Dense | OpenHands | Outcome | 13.6 | |
- | |
Overview of general foundation models evaluated on issue resolution. The table details the specific inference scaffolds (e.g., OpenHands, Agentless) employed during the evaluation process to achieve the reported results.
| Model Name | Size | Arch. | Inf. Scaffold | Reward | Res.(%) | Code | Model |
|---|---|---|---|---|---|---|---|
| KAT-Coder | - | - | Claude Code | Outcome | 73.4 | - |  |
| MiMo-V2-Flash | 309B-A15B | MoE | Agentless | Outcome | 73.4 | |
|
| Deepseek V3.2 | 671B-A37B | MoE | Claude Code, RooCode | - | 73.1 | |
|
| Kimi-K2-Instruct | 1T | MoE | Agentless | Outcome | 71.6 | - | |
| Qwen3-Coder | 480B-A35B | MoE | OpenHands | Outcome | 69.6 | |
|
| GLM-4.6 | 355B-A32B | MoE | OpenHands | Outcome | 68.0 | - | |
| gpt-oss-120b | 116.8B-A5.1B | MoE | Internal tool | Outcome | 62.0 | |
|
| Minimax M2 | 230B-10B | MoE | R2E-Gym | Outcome | 61.0 | |
|
| gpt-oss-20b | 20.9B-A3.6B | MoE | Internal tool | Outcome | 60.0 | |
|
| GLM-4.5-Air | 106B-A12B | MoE | OpenHands | Outcome | 57.6 | - | - |
| Minimax M1-80k | 456B-A45.9B | MoE | Agentless | Outcome | 56.0 | |
|
| Minimax M1-40k | 456B-A45.9B | MoE | Agentless | Outcome | 55.6 | |
|
| Seed1.5-Thinking | 200B-A20B | MoE | - | Outcome | 47.0 | |
- |
| Llama 4 Maverick | 400B-A17B | MoE | mini-SWE-agent | Outcome | 21.0 | |
|
| Llama 4 Scout | 109B-17B | MoE | mini-SWE-agent | Outcome | 9.1 | |
|
# First time: install dependencies
pip install flask flask-cors sqlalchemy pyyaml requests
# Full update + start admin server
# (refreshes news, re-renders README/docs, builds static site, then serves)
python start.py
# Or force re-import from YAML/CSV first
python start.py --initOpen http://localhost:5000/admin to manage papers, datasets, and methods.
| Command | Description |
|---|---|
python start.py |
Full update (news + render + build) then start server |
python start.py --init |
Re-import from YAML/CSV, then full update + start |
python start.py --no-update |
Start server without running update steps |
python start.py --port 8080 |
Use a custom port |
python start.py --news |
Refresh This Month's Papers only and exit |
python start.py --render |
Re-render README/docs from DB only and exit |
python start.py --build |
Build static site (mkdocs) only and exit |
We welcome contributions! To add new papers or tables:
- Fork this repository
- Add entries via the admin interface (
python start.py→localhost:5000/admin)
— or manually edit the YAML/CSV files indata/ - Run
python start.py --initif you edited files directly - Submit a PR with your changes
The application of LLMs in the programming domain has witnessed explosive growth. Early research focused primarily on function-level code generation, with benchmarks such as HumanEval serving as standard metrics. However, generic benchmarks often fail to capture the nuances of real-world development. To bridge this gap, recent initiatives have attempted to extend evaluation tasks to align more closely with realistic software development scenarios, revealing the limitations of general models in specialized domains. Concurrently, methods are also evolving to capture these broader contexts. While foundational approaches primarily relied on SFT or standard retrieval-augmented generation, RL-based methods emerged as a pivotal direction for handling complex coding tasks.
Related:
- HumanEval: Evaluating Large Language Models Trained on Code
- Program Synthesis: Program Synthesis with Large Language Models
- Repository-Level Code Completion: RLCoder: Reinforcement Learning for Repository-Level Code Completion
- Domain-Specific Benchmarks: Top General Performance = Top Domain Performance? DomainCodeBench
- Long-Context Code Models: Long Code Arena
- Multitask Fine-Tuning: MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning
- RAG for Code: RAG or Fine-tuning? A Comparative Study on LCMs-based Code Completion in Industry, Repoformer: Selective Retrieval for Repository-Level Code Completion, CodeRAG-Bench
- Code Generation Survey: A Survey on Large Language Models for Code Generation
The primary goal of this task is to autonomously construct complete and executable software systems starting from high-level natural language requirements. Unlike code completion, it necessitates covering the Software Development Life Cycle (SDLC), including requirement analysis, system design, coding, and testing. To address the complexity and potential logic inconsistencies in this process, state-of-the-art frameworks leverage multi-agent collaboration, simulating human development teams to decompose complex tasks into streamlined and verifiable workflows.
Related:
- ChatDev: Communicative Agents for Software Development
- MetaGPT: Meta Programming for Multi-Agent Collaborative Framework
- RPG: Repository Planning Graph for Unified and Scalable Codebase Generation
Issue resolution is intrinsically linked to the broader domain of automated software maintenance. Methodologies established in this field are frequently encapsulated as callable tools to augment the capabilities of LLMs in software development tasks.
Related:
- Bug Reproduction: AssertFlip, Automated Generation of Issue-Reproducing Tests
- Fault Localization:
- Code Search: A Benchmark for Localizing Code and Non-Code Issues
- Test Generation:
- Security: Is Vibe Coding Safe?
- Survey Papers:
Recent initiatives focus on automating the configuration of runtime environments for entire repositories. This capability develops in parallel with data construction for issue resolution.
Related:
- EnvBench: A Benchmark for Automated Environment Setup
- PIPer: On-Device Environment Setup via Online Reinforcement Learning
- Automated Benchmark Generation: Automated Benchmark Generation for Repository-Level Coding Tasks
Existing surveys primarily focus on code generation or other tasks within the software engineering domain. This paper bridges this gap by offering the first systematic survey dedicated to the entire spectrum of issue resolution, ranging from non-agent approaches to the latest agentic advancements.
Related:
- A Survey on Large Language Models for Code Generation
- Agents in software engineering: survey, landscape, and vision
- A Comprehensive Survey on Benchmarks and Solutions in Software Engineering of LLM-Empowered Agentic System
If you use this project or related survey in your research or system, please cite the following:
Li, Caihua, Guo, Lianghong, Wang, Yanlin, et al. (2026). Advances and Frontiers of LLM-based Issue Resolution in Software Engineering: A Comprehensive Survey. arXiv preprint arXiv:2601.11655.
BibTeX:
@article{li2026advances,
title={Advances and Frontiers of LLM-based Issue Resolution in Software Engineering: A Comprehensive Survey},
author={Li, Caihua and Guo, Lianghong and Wang, Yanlin and Guo, Daya and Tao, Wei and Shan, Zhenyu and Liu, Mingwei and Chen, Jiachi and Song, Haoyu and Tang, Duyu and Zhang, Hongyu and Zheng, Zibin},
journal={arXiv preprint arXiv:2601.11655},
year={2026},
eprint={2601.11655},
archivePrefix={arXiv},
primaryClass={cs.SE}
}We would like to express our sincere gratitude to:
-
The authors of cited papers who provided valuable feedback on how their work is presented in this survey, greatly improving its accuracy and comprehensiveness.
-
All contributors who have helped improve this project through issues, pull requests, and discussions.
-
The open-source community for developing the amazing tools and frameworks that made this project possible.
-
@chao-peng (Dr. Chao Peng), ByteDance Software Engineering Lab, for providing valuable suggestions on the Challenges and Opportunities section of our survey.
-
@EuniAI/awesome-code-agents for providing an excellent reference on managing survey papers through documentation systems and inspiring our project structure.
If you have any questions or suggestions, please contact us through:
- 📧 Email: noranotdor4@gmail.com
- 💬 GitHub Issues: Open an issue
This project is licensed under the MIT License - see the LICENSE file for details.
⭐ Star this repository if you find it helpful!
Made with ❤️ by the DeepSoftwareAnalytics team
Documentation | Paper | Tables | About | Cite