StateKV scales pretrained video VLMs linearly with video length at inference time without retraining

It reduces GFLOPs while preserving accuracy on VideoMME.