AI Paper Atlas

Research map / Video / Video-language understanding

VAST: Video Ability-Stratified Taxonomy for Data-Efficient Video Reasoning

Zhongan Wang, Xiaoyu Wen, Lingxiao Du, Kun Li, Zhiliang Wu, Xingcheng Xu, Qiaosheng Zhang, Chaochao Lu, Hehe Fan

CVPR 2026 · Video › Video-language understanding · approach: Video LLMs with fine-grained understanding

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add consistency reward (no SFT) (ablation)VideoVAST · Overall avg (16 frames)accuracy: 55.5 -> 55.8Improved
add consistency reward (no SFT) (ablation)VideoVAST · Overall avg (32 frames)accuracy: 55.8 -> 56.3Improved
add consistency reward (no SFT) (ablation)VideoVAST · Overall avg (64 frames)accuracy: 56.7 -> 56.7No clear effect
add consistency reward on SFT-initialized model (ablation)VideoVAST · Overall avg (32 frames)accuracy: 55.0 -> 56.2Improved
add consistency reward (ablation)VideoVAST · Video-Holmes (32 frames)accuracy: 37.7 -> 41.6Improved
add SFT initialization before RL (with CR) (ablation)VideoVAST · Overall avg (16 frames)accuracy: 55.8 -> 55.2Hurt
add SFT initialization before RL (with CR) (ablation)VideoVAST · Overall avg (32 frames)accuracy: 56.3 -> 56.2Hurt
add SFT initialization before RL (no CR) (ablation)VideoVAST · Overall avg (16 frames)accuracy: 55.5 -> 54.3Hurt
scale ability-stratified data from 1.5K to 15K (ablation)VideoVAST · Overall avg (16 frames)accuracy: 53.5 -> 55.8Improved
use ability-stratified VAST-15K instead of random 15K samples (ablation)VideoVAST · Overall avg (16 frames)accuracy: 47.6 -> 55.8Improved
use ability-stratified data instead of random on MMVU (ablation)VideoVAST · MMVU (mc)accuracy: 34.4 -> 66.1Improved
increase inference frames from 16 to 64 (RL with CR) (ablation)VideoVAST · Overall avgaccuracy: 55.8 -> 56.7Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy