Cowork-Gym: Synthesizing Interactive Work Environments at Scale
Abstract
Cowork-Gym synthesizes collaborative agent environments via iterative validation and trains Qwen models that substantially improve agentic benchmark performance.
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
Introducing Cowork-Gym, a framework for synthesizing executable work environments at scale to train agents for dynamic human–agent collaboration.
Real work involves changing requirements, user feedback, and edits to shared files. Cowork-Gym generates and validates environments through a build–execute–revise loop, while a stateful user simulator provides feedback and modifies the workspace—creating opportunities for agents to learn to clarify requirements, adapt plans, and revise their work.
- 🌍 Environments at scale: Combining taxonomy-guided and resource-grounded generation, we synthesize 64,888 environments spanning 14 application scenarios and 17 capability labels.
- 📈 Gains across model scales: Post-training Qwen models on these environments improves average scores across six agentic benchmarks by 8.7, 6.4, and 4.1 points for the 9B, 27B, and 35B-A3B models, respectively, over their corresponding base checkpoints.
- 🏆 Competitive performance: Our 27B model ranks second overall among the models evaluated in our study and achieves the highest average scores on PawBench and ClawEval, outperforming evaluated models with over 100B total parameters on both benchmarks.
QwenPaw-Flash models and our technical report are now available!
🤗 Models
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper