SLVMBench: leaderboard
SLVMBench (2026) embeds a tutorial video in a stream of unrelated videos and asks video LLMs 2,261 five-option questions that apply the demonstrated skill to a target video at a precise cutoff. Scores are accuracy in percent in the long-memory setting, with 10 minutes to 2 hours of distractor videos between the tutorial and the target; higher is better.
Metric: Accuracy (%) on 2,261 five-option questions on 1,220 videos in 11 categories, asked at a cutoff in a target video after a tutorial, with the tutorial followed by 10 minutes to 2 hours of unrelated distractor videos before the target (the long-memory setting); each model at its own frame sampling. Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 59.45 | #94 |
| 2 | GPT-5.2 | 44.89 | #176 |
| 3 | Qwen2.5-Omni-7B | 38.91 | #1789 |
| 4 | GPT-4o (2024-05-13) | 37.64 | #735 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=slvmbench · How It Works · Data refreshed daily, snapshot 2026-11-13.