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.
More from Multimodal
- evBaker standalone version in the works, focus on UI/UX — ssh4net · 2026-11-17
- Pixel art + Gaussian splats game experiments: all code, no assets, but roadblocks ahead — algo_diver · 2026-11-17
- Suno v6 music meets Seedance 2.5 video in a new AI demo — Kyrannio · 2026-11-17
- "The water forgot how to fall": surreal AI photo narrative thread — tallmetommy · 2026-11-17
- Recreating The Godfather's office scene with MiniMax H3 in ComfyUI: 3 shots, 37s — hucun16 · 2026-11-17
- Creator turns a Magnific Spaces retro-video workflow into an unhinged 1973 canned-bean commercial — charis_ai · 2026-11-17