HOT RADAR
llmHEAT 7.7°

AI CURATED HOT STORY · 11/16/2026

V-CAST Cuts Long-Video Inference Time by Half Without Training

1 reports archived1 independent sourcesupdated 11/16/2026, 22:07:00
Synthesis & Latest Updates
1 Sources Cross-Validated

Researchers from Sun Yat-sen University and Hong Kong Polytechnic University propose V-CAST, a training-free, plug-and-play visual token pruning method that allocates per-frame budgets and preserves original spatiotemporal coordinates before tokens enter the language model. On 768-frame audio-visual benchmarks it retains 97.1% of the uncompressed model's performance while cutting total inference time by 50.0%, and on Qwen3-VL-8B it keeps 98.2% average performance using only 25% of visual tokens.

LATEST/Researchers from Sun Yat-sen University and Hong Kong Polytechnic University propose V-CAST, a training-free, plug-and-play visual token pruning method that allocates per-frame budgets and preserves original spatiotemporal coordinates before tokens enter the language model. On 768-frame audio-visual benchmarks it retains 97.1% of the uncompressed model's performance while cutting total inference time by 50.0%, and on Qwen3-VL-8B it keeps 98.2% average performance using only 25% of visual tokens.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. 量子位 (微信公众号)T2·74 pts
    • V-CAST prunes visual tokens before they enter the language model, requiring no training and plugging in directly
    • On 768-frame audio-visual tests it retains 97.1% performance while cutting total inference time by 50.0%
    • It allocates frame budgets by adjacent-frame change and keeps original spatiotemporal coordinates to avoid position mismatch after merging