Synthetic cowork environments improve agents’ handling of changing shared workspaces

Cowork-Gym generated executable tasks with simulated user edits, producing consistent benchmark gains when used to post-train three Qwen models.

Chinese Tech
Anni Zou · Cheng Chen · Jiaji Deng · Weikang Zhou · Xinji Mai · Yue Wu · +15 more

Alibaba Token Foundry

Research Digest··2 min read
Zou et al.

The authors define a cowork environment as a task objective, initial workspace, available interfaces, user interaction policy and verification criteria.

Why this paper

From Alibaba Token Foundry

In one line

Cowork-Gym’s synthesized environments improve Qwen agent models by 4.1 to 8.7 points on six benchmarks.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ·No benchmark numbers found

Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.

§

Research Digest

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.

How we workSubscribe