Linear Scaling Video VLMs for Long Video Understanding

StateKV is an inference-only method for long-video vision-language models that keeps cross-frame context in a fixed-size recurrent state and uses a full cache for decoding.

Across three benchmarks and seven models, it stayed close to full self-attention and outperformed common sliding-window or recency-based streaming methods.

It reduces prefill cost to linear time without fine-tuning or architectural changes, making long-video understanding and streaming more scalable.