KiToke uses kernel-based global redundancy estimation and interval-aware token merging to compress video tokens in LLMs, outperforming prior training-free methods especially at retention ratios down to 1%.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
KiToke: Kernel-based Interval-aware Token Compression for Video Large Language Models
KiToke uses kernel-based global redundancy estimation and interval-aware token merging to compress video tokens in LLMs, outperforming prior training-free methods especially at retention ratios down to 1%.