MobileJudgeBench (A3 Final-State Judge): leaderboard
MobileJudgeBench (2026) measures how accurately an LLM judge decides from a recorded trajectory whether a mobile GUI agent completed its task, against human labels on 931 trajectories from six benchmarks. Accuracy in percent, higher is better, using the A3 judging method in final-state mode.
Metric: Judge accuracy (%; share of 931 human-annotated mobile-agent trajectories, from 6 benchmarks, 4 agent models and 68 apps, whose success or failure the judge classifies correctly; the A3 judge in final-state mode, temperature 0.7 as in its original implementation). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 83.8 | #141 |
| 2 | Qwen 2.5 VL 72B Instruct | 81.5 | #845 |
| 3 | Claude Sonnet 4.5 | 78.4 | #244 |
| 4 | GLM-4.6V | 78.1 | #698 |
| 5 | GPT-5 Mini | 76.4 | #349 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mobilejudgebench-a3-final-state-judge · How It Works · Data refreshed daily, snapshot 2026-11-12.