Research map / Large language models / Agents, code and math

RepoLaunch: Automating Build and Management of Code Repositories across Languages and Platforms

NeurIPS 2026 · Large language models › Agents, code and math · approach: Software engineering agents · 0 citations (OpenAlex)

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.

Introduces RepoLaunch, an LM-agent framework that automatically builds, tests, and manages code repositories across languages and OSes, and uses it to auto-create multi-language and Windows SWE-bench-Live task sets.

Problem

Building and testing software repositories at scale across diverse programming languages and operating systems is a manual, labor-intensive bottleneck that limits creation of executable SWE benchmark and training environments.

Approach

A multi-agent workflow (Preparation, Build, Management stages) where LM agents explore repos in Docker containers using language-specific base images and instructions, then distill rebuild commands and a structured test-log parser; applied to automatically create SWE tasks with validation from PR test status.

Key results

• RepoLaunch achieves 79.6% overall build success and 81.8% management-and-validation success on a 152-instance benchmark, versus 62.5% and 50.5% for SWE-agent.
• On the 152-instance benchmark RepoLaunch outperforms repo2run by 18% in build success, the prior Python/Linux-only system.
• The highest LM success rate on SWE-bench-Live/MultiLang and Windows tasks is 35.8% (GPT-5.5), with Java easiest and Rust hardest.

Surprising findings

• Simple repos resolvable with one command still encountered build or test errors in about one-third of cases, so rule-based shortcuts were not used.
• Failure to parse test logs is the dominant Management-stage failure across languages.

Limitations

• Multi-turn free agent consumes many tokens and is costly per instance.
• Rule-based and LM-based filtering of GitHub tasks may introduce bias.
• Timeout of 30 minutes causes failures on large resource-intensive repos; dataset size limited by compute.

Experimental setup

Datasets: SWE-bench-Verified SWE-bench-Multilingual SWE-bench-Pro SWE-bench-Live

Benchmarks: SWE-bench-Verified SWE-bench-Multilingual SWE-bench-Pro SWE-bench-Live/MultiLang SWE-bench-Live/Windows

Metrics: build success rate management success rate validation retention rate issue resolution success rate API cost per instance

Baselines: SWE-agent repo2run CXXCrafter pipreqs railpack

Builds on: SWE-bench SWE-agent repo2run

Closest papers by meaning

• RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing (ICML 2025 · similarity 0.90)
• SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents (NeurIPS 2025 · similarity 0.89)
• RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving (NeurIPS 2025 · similarity 0.89)
• Automated Benchmark Generation for Repository-Level Coding Tasks (ICML 2025 · similarity 0.88)
• GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging (AAAI 2026 · similarity 0.88)
• ProgramBench: Can Language Models Rebuild Programs From Scratch? (NeurIPS 2026 · similarity 0.88)