Research map / Large language models / Agents, code and math
RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RepoGraph, a line-level repository code graph whose ego-graph retrieval plugs into procedural (RAG, Agentless) and agent (AutoCodeRover, SWE-agent) frameworks to boost SWE-bench-Lite resolve rates and transfers to CrossCodeEval.
Problem
LLMs struggle with repository-level software engineering tasks such as resolving GitHub issues, since file-level retrieval and flat repository views miss cross-file dependencies and global structure.
Approach
Build a graph whose nodes are code lines (def/ref) and whose edges capture invoke and contain dependencies across the repository, then retrieve k-hop ego-graphs around a search term and feed them to existing procedural or agent pipelines as a plug-in.
Key results
- Agentless with GPT-4o plus RepoGraph reaches 29.67% resolve rate on SWE-bench-Lite (vs 27.33%), a state-of-the-art among open-source methods at submission.
- RAG with GPT-4 improves from 2.67% to 5.33% resolve rate with RepoGraph, roughly 99.6% relative gain.
- On CrossCodeEval with GPT-4o, RepoGraph raises code-match EM from 10.5 to 28.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add RepoGraph to RAG | RAG with GPT-4 · SWE-bench-Lite test | resolve rate: 2.67 -> 5.33 | Helped |
| Add RepoGraph to Agentless | Agentless with GPT-4o · SWE-bench-Lite test | resolve rate: 27.33 -> 29.67 | Helped |
| Add RepoGraph to AutoCodeRover | AutoCodeRover with GPT-4 · SWE-bench-Lite test | resolve rate: 19.00 -> 21.33 | Helped |
| Add RepoGraph to SWE-agent | SWE-agent with GPT-4o · SWE-bench-Lite test | resolve rate: 18.33 -> 20.33 | Helped |
| Add RepoGraph | Deepseek-Coder on CrossCodeEval Python · CrossCodeEval Python | Code Match EM: 10.2 -> 19.7 | Helped |
| Add RepoGraph | Deepseek-Coder on CrossCodeEval Python · CrossCodeEval Python | Identifier Match F1: 49.1 -> 58.9 | Helped |
| Add RepoGraph | GPT-4o on CrossCodeEval Python · CrossCodeEval Python | Code Match EM: 10.5 -> 28.7 | Helped |
| Add RepoGraph | GPT-4o on CrossCodeEval Python · CrossCodeEval Python | Identifier Match F1: 47.9 -> 61.3 | Helped |
| Summarize the 1-hop ego-graph instead of flattening it (ablation) | Agentless + RepoGraph · SWE-bench-Lite test | resolve rate: 29.67 -> 28.33 | Hurt |
| Use 2-hop retrieval and flatten the ego-graph (ablation) | Agentless + RepoGraph · SWE-bench-Lite test | resolve rate: 29.67 -> 26.00 | Hurt |
| Use 2-hop retrieval and summarize the ego-graph (ablation) | Agentless + RepoGraph · SWE-bench-Lite test | resolve rate: 29.67 -> 28.67 | Hurt |
Surprising findings
- Flattening the 2-hop ego-graph without summarization gives the lowest resolve rate (26.00%), below the baseline.
- Adding LLM summarization to the 1-hop ego-graph lowers resolve rate (28.33% vs 29.67%).
Limitations
- Only proprietary GPT-4 family and Claude models were evaluated; open-source models were not comprehensively tested.
- Experiments limited to SWE-bench-Lite and Python only.
- Agent frameworks incur substantially higher costs due to repeated search_repograph calls.
Experimental setup
Datasets: SWE-bench-Lite CrossCodeEval
Benchmarks: SWE-bench-Lite CrossCodeEval
Metrics: resolve rate patch application rate average cost average tokens exact match edit similarity identifier match F1 localization accuracy
Baselines: RAG Agentless AutoCodeRover SWE-agent
Builds on: Agentless SWE-agent AutoCodeRover RAG tree-sitter
Closest papers by meaning
- Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks (NeurIPS 2025 · similarity 0.89)
- RepoLaunch: Automating Build and Management of Code Repositories across Languages and Platforms (NeurIPS 2026 · similarity 0.88)
- NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents (ICML 2026 · similarity 0.88)
- RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving (NeurIPS 2025 · similarity 0.87)
- RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing (ICML 2025 · similarity 0.87)
- R2E: Turning any Github Repository into a Programming Agent Environment (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.