Video large language models (Video LLMs) are powerful but expensive to run because they process many visual tokens. Existing training-free compression methods rely on heuristic importance rules, which can over-represent large redundant visual groups and miss less frequent patterns, especially when the token budget is very small.
KiToke takes a different approach by treating visual tokens as samples from the video's spatiotemporal token distribution. It uses global embedding density as a proxy for representational redundancy and applies inverse-density sampling to allocate tokens across patterns of different frequencies. The method also aggregates unselected tokens into retained anchors within detected temporal intervals, reducing cross-transition mixing.
Across evaluated backbones, KiToke achieved the highest average accuracy among compared training-free methods and remained robust down to 1% token retention, where most baselines underperform uniform random sampling. On LLaVA-OneVision, it preserved 98.5% and 89.2% of the uncompressed model's four-benchmark mean accuracy at 10% and 1% retention, with 6.8x and 12.5x prefill speedups, respectively, including compression overhead.