GUI-RobustEval - Error Depth 1: leaderboard
Error recovery for computer-use agents: the agent takes over an Ubuntu desktop task from a prefix replayed from another agent's failed OSWorld trajectory and must finish it within 50 steps; GUI-RobustEval (Bu et al., Alibaba Cloud, ICML 2026). At error depth 1 the prefix runs one step past the expert-located root-cause error; success rate in percent, mean of 3 runs, higher is better.
Metric: Post-error success rate (%; the agent takes over a desktop task from an erroneous prefix replayed from another agent's failed OSWorld trajectory, error depth 1: one step past the root-cause error; up to 50 steps including the prefix; mean of 3 runs). Source: arxiv.org. Saturation forecast: Around August 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 38.7 | #982 |
| 2 | GPT-5.1 | 16.5 | #225 |
| 3 | Qwen 2.5 VL 7B Instruct | 3 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gui-robusteval-error-depth-1 · How It Works · Data refreshed daily, snapshot 2026-11-12.