Adobe Research and scientists from Stanford and Princeton propose video world models with long-term memory based on State-Space Models
Researchers from Adobe Research, Stanford University, and Princeton University have developed the LSSVWM architecture for video world models that addresses the problem of long-term memory. The model uses State-Space Models (SSMs) and a block-wise scanning scheme, as well as dense local attention, to combine efficiency with context retention. Experiments on the Memory Maze and Minecraft datasets showed significant improvement in information retention over long time intervals.
Princeton University (Sengupta Lab)
Synced27.07 · 17:06