OmniStarPro-FDQ (Online): leaderboard
Frame-level dense question answering on OmniStarPro (Yang et al., 2026): questions about fine-grained attributes or actions in the current frame arrive at dense intervals of a live stream, and the model times its own answers. A GPT-4o judge scores semantic correctness on a 0-10 scale; higher is better.
Metric: Semantic correctness (0-10; GPT-4o judge score of answers to frame-level dense questions asked at dense intervals about the current frame, the model timing its own answers). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LiveStarPro | 6.61 | |
| 2 | LiveStar | 6.44 | #1076 |
| 3 | MMDuet | 4.78 | |
| 4 | VideoLLM-online | 2.35 | |
| 5 | VideoLLM-MoD | 2.11 |
Interactive version: theaggregate.ai/benchmark?slug=omnistarpro-fdq-online · How It Works · Data refreshed daily, snapshot 2026-11-12.