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..
Sparse Diffusion Autoencoder for Test-time Adapting Prediction of Complex Systems
Authors: Jingwen Cheng, Ruikun Li, Huandong Wang, Yong Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations on representative systems demonstrate that Sparse Diff achieves an average prediction error reduction of 49.99% compared to baselines, requiring only 1% of the spatial resolution. Experimental evaluation on simulated and real-world systems demonstrates that Sparse Diff outperforms baselines by over 49.99% in long-term predictions, and reliably predicts full-space spatiotemporal dynamics using less than 1% of grid points as probes. In this section, we validate the accuracy and efficiency of Sparse Diff on simulated PDE systems and real-world datasets. Furthermore, we evaluate its robustness and generalization ability and examine the specific contribution of its components through ablation studies. |
| Researcher Affiliation | Academia | Jingwen Cheng Department of Electronic Engineering Tsinghua University Beijing, China Ruikun Li Shenzhen International Graduate School Tsinghua University Shenzhen, China Huandong Wang Department of Electronic Engineering BNRist, Tsinghua University Beijing, China Yong Li Department of Electronic Engineering BNRist, Tsinghua University Beijing, China |
| Pseudocode | No | The paper describes the methodology in prose and mathematical equations in sections 3 and 3.1, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is open-source: https://github.com/ tsinghua-fib-lab/Sparse Diff . |
| Open Datasets | Yes | Additionally, we also evaluate Sparse Diff s performance in real-world applications on an open-source climate record dataset [27]. We also evaluate Sparse Diff on the 2D wave equation with observations sampled on irregular meshes [47]. |
| Dataset Splits | Yes | We generate 50 training and 20 testing trajectories across varying flow conditions, with Reynolds numbers uniformly sampled: 10 training samples in the range Re [100, 500] and 10 out-of-distribution (OOD) samples in Re [500, 1000]. |
| Hardware Specification | No | The paper states in the NeurIPS Paper Checklist that hardware information can be found in a 'time cost section in the appendix', but no such section with specific hardware details (GPU models, CPU types, memory) is present in the provided document. The document mentions 'type of compute workers' but does not specify them. |
| Software Dependencies | No | Appendix B Model Architecture specifies 'ODE_method = 'rk4'' for numerical solver for ODE integration. However, it does not provide version numbers for any other key software components, libraries, or programming languages used (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | Appendix A provides parameters for the PDE systems (e.g., Lambda-Omega: 'µv and µv to 0.5, while β is 1.0'; Navier-Stokes: 'ν = 1.0, forcing amplitude A = 0.1, and phase shift s = 0'). It also mentions 'Temporal downsampling by a factor of 10 and 25 are applied to the LO and NS systems separately to reduce redundancy, while all the systems are spatially rescaled to a uniform 128 128 resolution.' Appendix B Model Architecture details hyperparameters such as 'hidden_dim = 1024', 'embedding_dim = 512', 'input_steps = 10', 'feature_dim = 256', 'num_heads = 8', 'dropout = 0.1'. |