RealStreamEval - Forward Active Responding: leaderboard

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 forward active responding tasks, where the model must wait until the evidence appears: repetition counting, sequential steps and clue-revealed responding). Source: arxiv.org. Saturation forecast: Around December 2027. 16 models tracked.

Top models

#ModelScore
1Qwen 3 VL 8B15.8
2Qwen 2 VL 7B12.6
3Qwen 2.5 VL 7B12.6

Interactive version: theaggregate.ai/benchmark?slug=realstreameval-forward-active-responding · How It Works · Data refreshed daily, snapshot 2026-09-26.