MEGA Hub

Cowork-Gym: Synthesizing Interactive Work Environments at Scale

Authors

Do you know Anni Zou?You can claim authorship or link another user.Do you know Cheng Chen?You can claim authorship or link another user.Do you know Jiaji Deng?You can claim authorship or link another user.Do you know Weikang Zhou?You can claim authorship or link another user.Do you know Xinji Mai?You can claim authorship or link another user.Do you know Yue Wu?You can claim authorship or link another user.Do you know Jianing Lin?You can claim authorship or link another user.Do you know Lipeng Xie?You can claim authorship or link another user.Do you know Lixiong Qin?You can claim authorship or link another user.Do you know Qingyue Zhao?You can claim authorship or link another user.Do you know Que Shen?You can claim authorship or link another user.Do you know Shuchang Tao?You can claim authorship or link another user.Do you know Xuchen Pan?You can claim authorship or link another user.Do you know Yaliang Li?You can claim authorship or link another user.Do you know Yunpeng Zhai?You can claim authorship or link another user.Do you know Yushuo Chen?You can claim authorship or link another user.Do you know Zhihai Wang?You can claim authorship or link another user.Do you know Zhuo Zhang?You can claim authorship or link another user.Do you know Zhaoyang Liu?You can claim authorship or link another user.Do you know Bolin Ding?You can claim authorship or link another user.Do you know Jingren Zhou?You can claim authorship or link another user.

Abstract

Cowork agents navigate dependencies across tools and files while adapting to changing user requirements and edits to shared workspaces. Scaling such environments requires maintaining consistency among task objectives, executable resources, user behavior, and success criteria. We introduce Cowork-Gym, a framework that synthesizes Cowork environments through a build-execute-revise loop that uses execution evidence and review feedback to validate and repair generated environments. To capture realistic Cowork dynamics, a stateful user simulator responds to the agent's actions through dialogue and direct edits to the shared workspace. We synthesize 64.9K environments across 14 application scenarios and use them to post-train 9B, 27B, and 35B-A3B Qwen models with supervised fine-tuning, reinforcement learning, and on-policy distillation. Averaged across six agentic benchmarks, these models outperform their corresponding base checkpoints by 8.7, 6.4, and 4.1 points, respectively. Ours-27B ranks second overall among the evaluated models and achieves the highest average scores on PawBench and ClawEval, outperforming models with over 100B total parameters on both benchmarks.

Community

00