KiToke: Distribution-Aware Token Compression for Video Large Language Models
KiToke is a training-free method designed to compress video tokens for Video LLMs to reduce high inference costs. It achieves this by estimating global token diversity using a kernel-based measure, allowing it to select tokens based on their overall importance across the entire video. It also preserves temporal structure by merging unselected tokens into retained anchors within detected temporal intervals, leading to better accuracy even with very aggressive token budgets.