KiToke Cuts 99% of Video Tokens While Retaining Accuracy

Researchers have released KiToke, a training-free token compression technique designed to drastically trim visual inputs for Video LLMs. Evaluated on LLaVA-OneVision, the method retains 89.2% of baseline uncompressed accuracy while discarding 99% of visual tokens. Operating at just a 1% retention ratio, KiToke outperformed previous state-of-the-art token reduction methods by 7 percentage points and achieved a 12.5x prefill speedup on an NVIDIA A100 GPU with minimal computational overhead.

Long video sequences often generate tens of thousands of visual tokens, creating quadratic compute and memory burdens for Transformer self-attention layers. KiToke solves this by selecting and merging projected visual tokens before language model prefill without altering model weights. The system utilizes kernel density estimation to evaluate token diversity across the entire video, assigning high-density scores to repetitive background frames or slow movements so they can be pruned. To prevent errors during scene transitions, content-adaptive temporal intervals ensure tokens are not improperly merged across cuts or semantic shifts.

Unlike earlier training-free approaches such as PruneVID, VidCom2, and FastVID, which struggle with dynamic motion, temporal redundancy, or missing rare details under extreme compression, KiToke treats token selection as a global diversity task. The framework is broadly applicable across model families and scales, demonstrating efficacy on LLaVA-OneVision, LLaVA-Video, and Qwen3-VL models ranging from 0.5B to 72B parameters. For AI practitioners, this approach enables significantly faster video inference and lower memory overhead without requiring costly retrainings or severe accuracy drops.