EVAThink Controlled 32 Frames - MVBench: leaderboard
EVAThink Controlled 32 Frames - MVBench: Video understanding accuracy (%). EVAThink arXiv 2611.13907v1 Table 1 controlled author reruns: 32 and 64 uniformly sampled frames, 256x28x28 pixels/frame. Original-protocol cited upper block is not ingested. Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct are explicit. Video-R1/VideoRFT/VideoAuto-R1 refer to released trained models; exact rows and frame budgets retained. EVAThink trained row is held until a released checkpoint can be located (paper claims availability, official repository currently exposes training code but no checkpoint link).
Metric: Video understanding accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 65.3 | #982 |
| 2 | Qwen 2.5 VL 7B Instruct | 60.5 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evathink-controlled-32-frames-mvbench · How It Works · Data refreshed daily, snapshot 2026-11-12.