V-CAST: training-free video token pruning keeps 98% accuracy with 25% tokens, cuts prefill 51%

量子位 · wechat · 2026-11-16

Researchers from Sun Yat-sen University, PolyU and others propose V-CAST, a training-free, plug-and-play video token pruning method for video LLMs. It allocates per-frame token budgets by local curvature of temporal feature change, selects tokens within frames via two complementary signals (deviation from frame mean + activation strength), and preserves original spatio-temporal coordinates. Results: 98.2% of Qwen3-VL-8B accuracy with 25% tokens at 32 frames; +1.5 pts over the best prior method at 256 frames; 97.1% retention with 50% less total inference time on 768-frame audio-video tasks; enables previously OOM 64-frame eval on Qwen3-VL-30B-A3B. Prefill latency drops 51.2%. The curvature allocator also boosts VisionZip and VidCom² when plugged in. Code is open-sourced.

Related event: V-CAST: Training-Free Visual Token Pruning Speeds Up Long Video Understanding by 51%(2 posts)→

Original post →

More from Multimodal

Multimodal channel →