Research map / Video / Video generation
Long-Context State-Space Video World Models
Ryan Po, Yotam Nitzan, Richard Zhang, Berlin Chen, Tri Dao, Eli Shechtman, Gordon Wetzstein, Xun Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add block-wise SSM scan instead of a single scan over all tokens (ablation) | Ours · Memory Maze reasoning | PSNR: 27.5 -> 28.2 | Improved |
| use layer-dependent block sizes instead of block size 1 everywhere (ablation) | Ours · Memory Maze reasoning | PSNR: 23.1 -> 28.2 | Improved |
| add long-context training with clean prefix frames (ablation) | Ours · Memory Maze reasoning | PSNR: 25.3 -> 28.2 | Improved |
| add frame local attention to Mamba2 (reasoning) | Mamba2 · Memory Maze reasoning | PSNR: 26.4 -> 27.5 | Improved |
| add frame local attention to Mamba2 (retrieval) | Mamba2 · Memory Maze retrieval | PSNR: 20.4 -> 19.3 | Hurt |
| add block-wise scan to Mamba2 + frame local attention (retrieval) | Mamba2 + Frame Local Attn · Memory Maze retrieval | PSNR: 19.3 -> 30.8 | Improved |
| use SSM + local attention instead of causal transformer limited to 192 frames | Causal transformer · Memory Maze retrieval | PSNR: 26.4 -> 30.8 | Improved |
| use full-context causal transformer instead of ours | Ours · Memory Maze retrieval | PSNR: 30.8 -> 32.6 | Improved |
| use SSM hybrid instead of DFoT on Minecraft | DFoT · TECO Minecraft reasoning | PSNR: 17.1 -> 17.8 | Improved |
| use SSM hybrid instead of full-context causal transformer on long generation | Causal (full context) · Memory Maze long generation | FVD: 45.1 -> 38.9 | Improved |
Closest papers by meaning
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.92)
- StateSpaceDiffuser: Bringing Long Context to Diffusion World Models (NeurIPS 2025 · similarity 0.91)
- WorldMem: Long-term Consistent World Simulation with Memory (NeurIPS 2025 · similarity 0.89)
- Addressable Memory for Video World Models (NeurIPS 2026 · similarity 0.88)
- SAMPO: Scale-wise Autoregression with Motion Prompt for Generative World Models (NeurIPS 2025 · similarity 0.88)
- Slot State Space Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
