Research map / Reinforcement learning / Visual and representation RL
SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents
Danlong Yuan, Wei Wu, Zhengren Wang, Xueliang Zhao, Huishuai Zhang, Dongyan Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · RL rollout, 3B | env prepare time (s): 88.86 -> 23.62 | Improved |
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · RL rollout, 3B | avg rollout time (s): 367.33 -> 272.71 | Improved |
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · SWE-bench Verified, 3B after RL | resolve rate (%): 8.6 -> 9.2 | No clear effect |
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · SWE-bench Verified, 7B after RL | resolve rate (%): 12.4 -> 11.8 | No clear effect |
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · SWE-bench Verified, SWE-Agent-7B after RL | resolve rate (%): 16.4 -> 16.8 | No clear effect |
| use MiniSandbox instead of image-based storage | SWE-MiniSandbox · SWE-smith env cache | environment storage: 295 GB -> 13.5 GB | Improved |
| use MiniSandbox instead of one-to-one images | SWE-MiniSandbox · SWE-bench Verified env cache | environment storage: 605 GB -> 89 GB | Improved |
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · reward computation | reward time (s): 14.89 -> 9.96 | Improved |
| use MiniSandbox instead of Docker containers | SWE-MiniSandbox · environment communication timeouts | timeouts per instance: 0.93 -> 0.39 | Improved |
| scale rollout count to 16 (256 parallel envs, 1 node) (ablation) | SWE-MiniSandbox · MiniSandbox vs Docker | env prepare time (s): MiniSandbox 20.09 -> 116.25 vs Docker 87.45 -> 123.79 | Hurt |
| scale from 1 to 2 nodes at 256 envs | SWE-MiniSandbox · MiniSandbox | env prepare time (s): 116.25 -> 20.72 | Improved |
| scale from 1 to 2 nodes at 256 envs | SWE-MiniSandbox · Docker | env prepare time (s): 123.79 -> 113.67 | No clear effect |
Closest papers by meaning
- SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents (NeurIPS 2025 · similarity 0.88)
- Toward Training Superintelligent Software Agents through Self-Play SWE-RL (ICML 2026 · similarity 0.88)
- Kimi-Dev: Agentless Training as Skill Prior for SWE-agents (ICLR 2026 · similarity 0.87)
- DOCKSMITH: Scaling Reliable Coding Environments via an Agentic Docker Builder (ICML 2026 · similarity 0.87)
- NAVIX: Scaling MiniGrid Environments with JAX (NeurIPS 2025 · similarity 0.86)
- JaxMARL: Multi-Agent RL Environments and Algorithms in JAX (NeurIPS 2024 · similarity 0.86)
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.
