Research map / Architectures and efficiency / Attention and transformers
Composition of Memory Experts for Diffusion World Models
Sebastian Stapf, Pablo Acuaviva, Aram Davtyan, Paolo Favaro
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 |
|---|---|---|---|
| use contrastive product of experts instead of plain PoE when combining all experts (ablation) | CoME (all experts) · Memory Maze | LPIPS: 0.192 -> 0.097 | Improved |
| add contrastive weighting to STM+LTM (ablation) | CoME · Memory Maze | LPIPS: 0.170 -> 0.114 | Improved |
| add contrastive baseline to the base model alone (ablation) | Base DiT · Memory Maze | LPIPS: 0.203 -> 0.200 | No clear effect |
| add long-term memory via LoRA test-time finetuning (ablation) | base DiT · Memory Maze | LPIPS: 0.203 -> 0.171 | Improved |
| combine STM and LTM experts (ablation) | CoME · Memory Maze | LPIPS: 0.171 -> 0.114 | Improved |
| increase LTM context frames from 50 to 450 (ablation) | LTM (LoRA rank 256) · Memory Maze | LPIPS: 0.162 -> 0.118 | Improved |
| use LoRA instead of full finetuning for LTM at low context (ablation) | LTM · Memory Maze (50 frames) | LPIPS: 0.221 (full) -> 0.162 (rank 256) | Improved |
| add STM or LTM memory to NWM planner | NWM · RECON planning | ATE: 1.13 -> 0.96 (CoME) | Improved |
| add memory experts to Oasis-style Minecraft world model | base diffusion WM · Minecraft Marsh | PSNR: 16.19 -> 17.78 | Improved |
| add Langevin correction steps during sampling (ablation) | CoME · Memory Maze | generation quality | No clear effect |
| increase LoRA rank for LTM memorization (compute cost) (ablation) | LTM · sampling overhead | sampling time overhead: r=64 about 4x vs r=16 about 2x | Hurt |
Closest papers by meaning
- StateSpaceDiffuser: Bringing Long Context to Diffusion World Models (NeurIPS 2025 · similarity 0.89)
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.89)
- EDELINE: Enhancing Memory in Diffusion-based World Models via Linear-Time Sequence Modeling (NeurIPS 2025 · similarity 0.89)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · similarity 0.87)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.87)
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.86)
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.
