VenusBench-Mobile - Vague Instructions: leaderboard

Task completion of mobile GUI agents on VenusBench-Mobile (2026), user-intent Android tasks in emulated apps with screenshot observations and task-specific verification, on the vague instructions tasks (underspecified requests such as an omitted app name). Share of tasks completed, in percent; higher is better.

Metric: Task success rate (%) on the 16 vague-instruction (underspecified requests such as an omitted app name) tasks on VenusBench-Mobile, user-intent-driven Android tasks in emulated apps (screenshot-only observations, step-by-step GUI actions, task-specific verification functions with an MLLM checker where visual or semantic judgment is needed); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1MAI-UI-8B25
2UI-Venus-7B25
3Qwen 3 VL 8B18.8#917
4Qwen 3 VL 30B A3B18.8#784
5GUI-Owl-7B18.8
6UI-Venus-72B12.5
7Mobile-Agent-v3 + GUI-Owl-7B6.3

Interactive version: theaggregate.ai/benchmark?slug=venusbench-mobile-vague-instructions · How It Works · Data refreshed daily, snapshot 2026-11-21.