KeyRec: Bounded Visual Memory for Streaming and Long-Video Understanding
KeyRec keeps video memory bounded by splitting recent detail from older event evidence.
The paper describes a training-free framework for long-video and streaming VLM use, aimed at avoiding the cost of dense visual tokens piling up over time. It stores fine-grained recent observations in a cache while maintaining older evidence in an event bank updated through add, merge, and evict steps. At question time, a text-only router divides a fixed readout budget between those two memories without reprocessing historical frames. The authors report best compressed performance in 13 of 15 benchmark settings using 10% of the dense visual-token budget. Source: HF Daily Papers' note
The paper describes a training-free framework for long-video and streaming VLM use, aimed at avoiding the cost of dense visual tokens piling up over time. It stores fine-grained recent observations in a cache while maintaining older evidence in an event bank updated through add, merge, and evict steps. At question time, a text-only router divides a fixed readout budget between those two memories without reprocessing historical frames. The authors report best compressed performance in 13 of 15 benchmark settings using 10% of the dense visual-token budget. Source: HF Daily Papers' note
score 5