RealStreamEval - Backward Tracing: leaderboard
Streaming video question answering under the RealStreamEval protocol (Wen et al., 2026): OVO-Bench clips arrive at 0.5 frames per second and at every step the model decides whether to answer or stay silent. Mean of three backward-tracing tasks about earlier parts of the stream (episodic memory, action sequence and hallucination detection), judged by Qwen3-VL-235B-A22B-Instruct with a penalty for repeated answers, in percent; higher is better.
Metric: Score (%; RealStreamEval on OVO-Bench tasks and questions: frames arrive at 0.5 fps and the model decides at every step whether to answer or stay silent; answers judged by Qwen3-VL-235B-A22B-Instruct with timing windows and a verbosity or repetition penalty; 768 visual tokens per frame; mean of the three backward-tracing tasks: episodic memory, action sequence and hallucination detection). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 7B | 49.8 | #2273 |
| 2 | Qwen 2.5 VL 7B | 46.9 | #1796 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=realstreameval-backward-tracing · How It Works · Data refreshed daily, snapshot 2026-11-12.