Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
StateSpaceDiffuser: Bringing Long Context to Diffusion World Models
Authors: Nedko Savov, Naser Kazemi, Deheng Zhang, Danda Pani Paudel, Xi Wang, Luc V Gool
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments show that State Space Diffuser significantly outperforms a strong diffusion-only baseline, maintaining a coherent visual context for an order of magnitude more steps. It delivers consistent views in both a 2D maze navigation and a complex 3D environment. |
| Researcher Affiliation | Academia | 1 INSAIT, Sofia University "St. Kliment Ohridski" 2 ETH Zurich 3 TU Munich |
| Pseudocode | No | The paper includes architectural diagrams (e.g., Figure 3), but no explicit pseudocode blocks or algorithms. |
| Open Source Code | No | Project page: https://insait-institute.github.io/State Space Diffuser/. Answer: [Yes] Justification: Data generation and model code will be provided |
| Open Datasets | Yes | We evaluate on two environments. (1) A simpler 2D maze environment (Mini Grid)... We create a dataset based on the partially observed Mini Grid maze environment [12]. And (2) a 3D first-person shooter game (CSGO) [60]. |
| Dataset Splits | No | We train and test on the same set of 34 samples. The paper mentions using "Mini Grid test set" and discusses training for a certain number of iterations, but it does not specify explicit train/validation/test splits (e.g., percentages or sample counts) for the main datasets. |
| Hardware Specification | Yes | We use 8 A100 GPUs for all models except for the Mini Grid State Space World Model, which was trained on 4 A100 GPUs for Mini Grid models. |
| Software Dependencies | No | The paper mentions general software components like "Adam optimizer" and diffusion models but does not provide specific version numbers for any libraries, frameworks, or programming languages used. |
| Experiment Setup | Yes | All models use the Adam optimizer. The State Space World Model is trained with a learning rate of 5e-5 and batch size 136 (Mini Grid) or batch size 272 (CSGO). Both the Mini Grid and CSGO models are trained for 70k iterations on sequence size 16. State Space Diffuser includes 600M parameters, and is trained with a learning rate 1e-4, weight decay 1e-2, grad norm clip 10 and batch size 64. The Mini Grid model is trained for 77k iterations, CSGO 220k iterations. |