OmniStarPro-RNG (Online): leaderboard
Real-time narration generation on OmniStarPro (Yang et al., 2026) in the fully online setting, where the model watches the stream and decides by itself when to speak or stay silent. A GPT-4o judge scores the first response in each ground-truth clip for semantic accuracy, language quality and completeness on a 0-10 scale; higher is better.
Metric: Semantic correctness (0-10; GPT-4o judge score of the earliest response in each ground-truth clip, mean of semantic accuracy, language quality and information completeness; the model decides by itself when to speak, 1,000 evaluation streams of the OmniStarPro-Live partition). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LiveStarPro | 3.27 | |
| 2 | LiveStar | 3.19 | #1076 |
| 3 | MMDuet | 1.93 | |
| 4 | VideoLLM-online | 1.68 | |
| 5 | VideoLLM-MoD | 1.66 |
Interactive version: theaggregate.ai/benchmark?slug=omnistarpro-rng-online · How It Works · Data refreshed daily, snapshot 2026-11-12.