OmniStarPro-Long - TBR (Medium Span): leaderboard
Temporal backtracking: the model must locate the most recent past occurrence of a queried event and report its timestamp or an attribute, on the OmniStarPro-Long partition of OmniStarPro (Yang et al., 2026), streams of 10 minutes to over an hour. Accuracy in percent judged by GPT-4o, on queries whose evidence lies 10 to 30 minutes before the query; higher is better.
Metric: Recall accuracy (%; share of temporal backtracking (TBR) queries, which ask for the timestamp or an attribute of the most recent past occurrence of a queried event, that a GPT-4o judge marks as matching the ground truth; evidence 10-30 minutes before the query; OmniStarPro-Long partition, streams of 10 minutes to over an hour). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LiveStarPro | 46.3 | |
| 2 | LiveStar | 30.2 | #1076 |
| 3 | MMDuet | 21.3 | |
| 4 | VideoLLM-MoD | 16.7 | |
| 5 | VideoLLM-online | 16 |
Interactive version: theaggregate.ai/benchmark?slug=omnistarpro-long-tbr-medium-span · How It Works · Data refreshed daily, snapshot 2026-11-12.