PriMobiBench - User Profile Construction (Iterative): leaderboard
PriMobiBench (2026) user profile construction under iterative prompting, where the model refines the profile over several rounds after reading the screenshots of a mobile agent trajectory. Share of the ground-truth profile attributes inferred correctly, in percent, judged by GPT-5-mini; higher means more is extracted.
Metric: Profile success rate (%; share of the ground-truth attributes of 50 synthetic user profiles that the model correctly infers after reading every screenshot of a MobiLeak trajectory (320 trajectories across 16 Android apps with 2,960 inserted privacy elements in 25 attribute types), judged for semantic entailment by GPT-5-mini; iterative prompting: the profile is refined over several rounds in the same dialogue). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 71.2 | #349 |
| 2 | GLM-4.5V | 62.1 | #739 |
| 3 | Gemma 3 27B | 49.8 | #1638 |
| 4 | Qwen 2.5 VL 7B | 41.2 | #1796 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=primobibench-user-profile-construction-iterative · How It Works · Data refreshed daily, snapshot 2026-11-12.