Cowork-Gym: Synthesizing Interactive Work Environments at Scale
Cowork-Gym post-trains Qwen agents on 64.9K synthetic workplaces, lifting six-benchmark scores.
Cowork-Gym synthesizes interactive agent work environments with a build-execute-revise loop and a stateful user simulator that edits shared workspaces. The authors generate 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 over six agentic benchmarks, the models improve 8.7, 6.4, and 4.1 points over their bases; Ours-27B ranks second overall and tops PawBench and ClawEval, beating models above 100B parameters on both.
- Cowork-Gym synthesizes 64.9K interactive work environments across 14 scenarios.
- A stateful user simulator edits shared workspaces and replies in dialogue.
- Post-trained 9B, 27B, and 35B-A3B Qwen models gain 8.7, 6.4, and 4.1 points.
- Ours-27B leads PawBench and ClawEval, beating models over 100B parameters.
- AI models
- Qwen-9BQwen-27BQwen-35B-A3B
From the original article
1 min read at huggingface.co
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.
By Anni Zou, Cheng Chen, Jiaji Deng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.