Back

OSWorld V2.1 Research Atlas (Chinese)

2026-10-19
Computer UseBenchmarkOSWorldEvaluation

An interactive Chinese research atlas I built for OSWorld V2.1. It is a standalone page.

Open the atlas →

Built against osworld-v2.1 (released 2026-09-16), it covers:

  • Chinese requirements, deliverables, workflows, grading, estimated difficulty and atomic-skill breakdowns for all 108 tasks, with real trajectory screenshots
  • A per-task score matrix for five v2.1 runs (Sai, Darwinx, GPT-6 Astra, Muse Spark 1.3 / 1.2), plus trajectory-based failure analysis of the 20 hardest tasks
  • How each of the ten official challenge tags is set up, graded and failed by frontier models
  • Statistics and training priorities for 54 atomic skills
  • The environment and evaluation pipeline: task classes, assets, self-hosted websites, images, setup, user simulator, grading and release pinning

Text versions (Chinese): task catalog · environment pipeline · atomic skills · paper notes

Back
···