V-CAST: Training-Free Visual Token Pruning Speeds Up Long Video Understanding by 51%

Researchers from Sun Yat-sen University and PolyU proposed V-CAST, a training-free, plug-and-play spatiotemporal token pruning method that keeps about 97-98% of long video understanding performance with only 25% of visual tokens, speeding up prefill by 51%.

2026-11-16 ~ 2026-11-17 · 2 related posts

1 near-duplicate retellings: 量子位