Adobe Research and Stanford and Princeton Scientists Propose Video World Models with Long-Term Memory Based on State-Space Models

A new paper, "Long-Context State-Space Video World Models" by researchers from Stanford University, Princeton University, and Adobe Research, proposes a solution to the long-term memory bottleneck in video world models. The core problem is that traditional attention layers have quadratic computational complexity with respect to sequence length, making long-term memory impractical. The authors introduce the Long-Context State-Space Video World Model (LSSVWM), which leverages State-Space Models (SSMs) for efficient causal sequence modeling. Key design choices include a block-wise SSM scanning scheme that trades off some spatial consistency for extended temporal memory, and dense local attention to maintain fine-grained details. Two training strategies, Diffusion Forcing and Frame Local Attention, further improve long-context performance. Evaluations on Memory Maze and Minecraft datasets show that LSSVWM substantially surpasses baselines in preserving long-range memory while maintaining practical inference speeds.