KnowU-Bench - Proactive (Hard): leaderboard
Proactive mobile assistance on KnowU-Bench (2026), hard split: 32 situations at the higher of the two difficulty levels with no instruction, where the agent must decide whether to act, seek consent or stay silent and must respect a rejection by the simulated user. Success rate in percent; higher is better.
Metric: Success rate (%) on the 32 hard proactive tasks (no instruction: act, ask for consent or stay silent, and respect a rejection) in the reproducible KnowU-Bench Android emulator, noisy full-history memory setting (the complete user logs plus irrelevant history), gpt-4o as the user simulator grounded in a hidden user profile, scored by rule-based checks fused with a rubric-conditioned LLM judge; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 65.6 | #828 |
| 2 | Seed 2.0 Pro | 62.5 | #222 |
| 3 | Qwen 3.5 397B A17B | 56.3 | #264 |
| 4 | Claude Sonnet 4.6 | 53.1 | #173 |
| 5 | Qwen 3 VL 32B | 34.4 | #622 |
| 6 | UI-Venus-1.5-8B | 31.3 | |
| 7 | MAI-UI-8B | 22.2 | |
| 8 | Qwen 3 VL 8B | 21.9 | #917 |
| 9 | GUI-OWL-1.5-8B | 21.9 | |
| 10 | Qwen 3.5 122B A10B | 12.5 | #346 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=knowu-bench-proactive-hard · How It Works · Data refreshed daily, snapshot 2026-11-21.