HyperTrack: leaderboard
HyperTrack (Qu et al., Wuhan University and Xiaomi, ICML 2026) tests mobile GUI agents offline on real tasks in Chinese Android apps: at each step of a reference trajectory the model sees the screen and the task history and predicts the next action, using its recommended prompt in the authors' GUIEvalKit toolkit. A step counts only when the action type and its parameters (click point, scroll direction or typed text) both match the reference; the score is step exact match in percent, with OPEN steps excluded; higher is better.
Metric: Step Exact Match (%; predicted action type and parameters both correct, offline step-by-step evaluation on the reference trajectory, OPEN steps excluded). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-4.5V | 72.97 | #739 |
| 2 | GLM-4.1V-9B (Thinking) | 67.14 | #1061 (GLM-4.1V-9B) |
| 3 | Qwen 3 VL 8B Instruct | 66.28 | #982 |
| 4 | Qwen 3 VL 4B Instruct | 65.81 | #1281 |
| 5 | Qwen 2.5 VL 7B Instruct | 64.17 | #1842 |
| 6 | Qwen 3 VL 8B (Thinking) | 59.35 | #887 |
| 7 | Qwen 3 VL 4B (Thinking) | 56.96 | #1303 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hypertrack · How It Works · Data refreshed daily, snapshot 2026-11-12.