Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SWE-World: Building Software Engineering Agents in Docker-Free Environments

Shuang Sun, Huatong Song, Lisheng Huang, Jinhao Jiang, Ran Le, Zhihao Lv, Zongchao Chen, Yiwen Hu, Wenyang Luo, Wayne Xin Zhao, Yang Song, Hongteng Xu, Tao Zhang, Ji-Rong Wen

TL;DR

The paper tackles the resource bottlenecks of dependency-heavy SWE pipelines by introducing SWE-World, a Docker-free surrogate environment that combines a deterministic Sandbox with LLM-based models (SWT and SWR) to simulate execution feedback and test outcomes. The approach enables fully Docker-free training (SFT and RL) and supports test-time scaling, while leveraging open-source SWE data and a CoT-backed distillation strategy to improve reasoning fidelity. Key contributions include the Docker-free environment, effective agent training without physical execution, scalable data utilization, and strong TTS performance that matches or surpasses Docker-based pipelines. The results demonstrate substantial infrastructure savings and practical potential for large-scale SWE research and iteration.

Abstract

Recent advances in large language models (LLMs) have enabled software engineering agents to tackle complex code modification tasks. Most existing approaches rely on execution feedback from containerized environments, which require dependency-complete setup and physical execution of programs and tests. While effective, this paradigm is resource-intensive and difficult to maintain, substantially complicating agent training and limiting scalability. We propose SWE-World, a Docker-free framework that replaces physical execution environments with a learned surrogate for training and evaluating software engineering agents. SWE-World leverages LLM-based models trained on real agent-environment interaction data to predict intermediate execution outcomes and final test feedback, enabling agents to learn without interacting with physical containerized environments. This design preserves the standard agent-environment interaction loop while eliminating the need for costly environment construction and maintenance during agent optimization and evaluation. Furthermore, because SWE-World can simulate the final evaluation outcomes of candidate trajectories without real submission, it enables selecting the best solution among multiple test-time attempts, thereby facilitating effective test-time scaling (TTS) in software engineering tasks. Experiments on SWE-bench Verified demonstrate that SWE-World raises Qwen2.5-Coder-32B from 6.2\% to 52.0\% via Docker-free SFT, 55.0\% with Docker-free RL, and 68.2\% with further TTS. The code is available at https://github.com/RUCAIBox/SWE-World

SWE-World: Building Software Engineering Agents in Docker-Free Environments

TL;DR

The paper tackles the resource bottlenecks of dependency-heavy SWE pipelines by introducing SWE-World, a Docker-free surrogate environment that combines a deterministic Sandbox with LLM-based models (SWT and SWR) to simulate execution feedback and test outcomes. The approach enables fully Docker-free training (SFT and RL) and supports test-time scaling, while leveraging open-source SWE data and a CoT-backed distillation strategy to improve reasoning fidelity. Key contributions include the Docker-free environment, effective agent training without physical execution, scalable data utilization, and strong TTS performance that matches or surpasses Docker-based pipelines. The results demonstrate substantial infrastructure savings and practical potential for large-scale SWE research and iteration.

Abstract

Recent advances in large language models (LLMs) have enabled software engineering agents to tackle complex code modification tasks. Most existing approaches rely on execution feedback from containerized environments, which require dependency-complete setup and physical execution of programs and tests. While effective, this paradigm is resource-intensive and difficult to maintain, substantially complicating agent training and limiting scalability. We propose SWE-World, a Docker-free framework that replaces physical execution environments with a learned surrogate for training and evaluating software engineering agents. SWE-World leverages LLM-based models trained on real agent-environment interaction data to predict intermediate execution outcomes and final test feedback, enabling agents to learn without interacting with physical containerized environments. This design preserves the standard agent-environment interaction loop while eliminating the need for costly environment construction and maintenance during agent optimization and evaluation. Furthermore, because SWE-World can simulate the final evaluation outcomes of candidate trajectories without real submission, it enables selecting the best solution among multiple test-time attempts, thereby facilitating effective test-time scaling (TTS) in software engineering tasks. Experiments on SWE-bench Verified demonstrate that SWE-World raises Qwen2.5-Coder-32B from 6.2\% to 52.0\% via Docker-free SFT, 55.0\% with Docker-free RL, and 68.2\% with further TTS. The code is available at https://github.com/RUCAIBox/SWE-World
Paper Structure (41 sections, 13 equations, 4 figures, 11 tables)

This paper contains 41 sections, 13 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Execution Environments for SWE
  5. From Docker-Based to LLM-Based Environments
  6. SWE-World: LLM-Based Docker-Free Environment
  7. System Architecture
  8. Lightweight Sandbox for Navigation and Editing
  9. SWT: Step-Level Execution Feedback Simulation
  10. SWR: Test Report and Reward Generation for Evaluation Simulation
  11. Training SWT and SWR
  12. Data Collection from Real Docker Rollouts
  13. Reverse-Reasoning Distillation for CoT Backfilling
  14. Training SWE Agents with SWE-World
  15. Software Eengineering Data Preparation
  16. ...and 26 more sections

Figures (4)

  • Figure 1: Overview of SWE-World. Left: We collect agent-Docker interaction data to train the SWE-World Transition Model (SWT) and SWE-World Reward Model (SWR). Middle: SWE-World forms a Docker-free surrogate environment, enabling scalable agent enhancement via SFT, RL, and Test-Time Scaling. Right: Comparison of Code Agent trajectories generated based on Docker and SWE-World.
  • Figure 2: RL training dynamics using SWT-32B. Orange lines denote average reward; Green dashed lines denote mean interaction turns. Left: Main experiment using CoT-enhanced SWR-32B shows stable learning. Right: Comparison with non-CoT SWR-32B leads to trajectory length collapse, indicating reward hacking.
  • Figure 3: Test-time scaling on SWE-bench Verified: comparing SWR-32B with prior verifiers.
  • Figure 4: Qualitative fidelity of SWE-World simulation.