The Year in AI Papers
The year/Qwen

Paper 2602.02361

SWE-Universe: Scale Real-World Verifiable Environments to Millions

Mouxiang Chen, Lei Zhang, Yunlong Feng, Xuwu Wang, Wenting Zhao, Ruisheng Cao, Jiaxi Yang, Jiawei Chen, Mingze Li, Zeyao Ma, Hao Ge, Zongmeng Zhang, et al.

Published
Feb 2026
Research lab
Qwen
Citations
7
GitHub
Not linked

01 In brief

Summary

SWE-Universe is a framework for automatically constructing real-world software engineering (SWE) verifiable environments from GitHub pull requests (PRs) at a million scale.

It addresses challenges of low production yield, weak verifiers, and prohibitive cost using a building agent powered by a custom-trained MoE model (Qwen-Next-80B-A3B).

The agent uses iterative self-verification and in-loop hacking detection to ensure high-fidelity, executable verifiers.

The framework produced 807,693 multilingual instances from 52,960 repositories, surpassing existing datasets.

Mid-training on 500K successful trajectories (30B tokens) improved Qwen3-Next-80A3's SWE-Bench Verified score from 50.3% to over 61% and SWE-Bench Multilingual from ~31% to over 46%.

Agentic reinforcement learning on Qwen3-30B-A3B improved SWE-Bench Multilingual from ~32% to 42.0%.

Applied to Qwen3-Max-Thinking, the method achieved 75.3% on SWE-Bench Verified.

The model achieved a 78.44% non-hacked success rate on a custom benchmark, surpassing Claude-Opus-4.5 (77.81%).

02 From the paper

Abstract

We propose SWE-Universe, a scalable and efficient framework for automatically constructing real-world software engineering (SWE) verifiable environments from GitHub pull requests (PRs). To overcome the prevalent challenges of automatic building, such as low production yield, weak verifiers, and prohibitive cost, our framework utilizes a building agent powered by an efficient custom-trained model. This agent employs iterative self-verification and in-loop hacking detection to ensure the reliable generation of high-fidelity, verifiable tasks. Using this method, we scale the number of real-world multilingual SWE environments to a million scale (807,693). We demonstrate the profound value of our environments through large-scale agentic mid-training and reinforcement learning. Finally, we applied this technique to Qwen3-Max-Thinking and achieved a score of 75.3% on SWE-Bench Verified. Our work provides both a critical resource and a robust methodology to advance the next generation of coding agents.