VSAS-Bench (Synchronous) - Cumulative: leaderboard

VSAS-Bench (2026) Cumulative tasks, which require recalling and reasoning over past events in a streamed video, under the synchronous protocol; per-timestep responses judged against dense reference annotations, accuracy (%), higher is better.

Metric: Mean average accuracy (%) on Cumulative tasks (recalling and reasoning over past events) under the synchronous protocol (the model runs in lockstep with the stream, so inference latency does not count; video VLMs adapted with the sliding-window memory policy), over VSAS-Bench's 92 videos with dense 1 FPS free-form annotations; each timestep's response is judged against the annotation by GPT-5 (medium reasoning) on a 0-1 scale, shown as %; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5Medium65.3#178
2Qwen 3 VL 32B Instruct58.3#606
3InternVL3-38B53.5#939
4Qwen 3 VL 8B Instruct52.3#1029
5Qwen 3 VL 4B Instruct47.9#1332
6InternVL3-8B41.8#1782
7Qwen 3 VL 2B Instruct38.3#2367
8Dispider33.5#3815
9StreamBridge [VSAS-Bench checkpoint unspecified]26.8
10InternVL3-2B23.3#3290
11Flash-VStream-7B5.6
12VideoLLM-Online1.6

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=vsas-bench-synchronous-cumulative · How It Works · Data refreshed daily, snapshot 2026-11-21.