Research map / Architectures and efficiency / Pruning and efficient tuning
Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion
Ruikun Li, Huandong Wang, Jingtao Ding, Yuan Yuan, Qingmin Liao, Yong Li
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 |
|---|---|---|---|
| generate full expert weights with weight-space diffusion (DynaDiff) instead of meta-learning modulation | DynaDiff · Cylinder Flow out-domain | RMSE: 0.082 (GEPS) -> 0.065 | Improved |
| use 1M-parameter generated experts instead of ~500M foundation models | DynaDiff · Navier-Stokes out-domain | RMSE: 0.074 (FNO) -> 0.064 | Improved |
| use DynaDiff weight graph, functional loss and prompter instead of D2NWG latent weight diffusion | weight-space generation · Lambda-Omega out-domain | RMSE: 0.105 -> 0.091 | Improved |
| generate weights with CVAE instead of diffusion | weight-space generation · Kolmogorov Flow | RMSE: 4.454 in-domain, 29.089 out-domain | Hurt |
| add function loss to VAE training (ablation) | DynaDiff · Kolmogorov Flow and Navier-Stokes | prediction accuracy | Improved |
| add domain-adaptive initialization when building the model zoo (ablation) | DynaDiff · Kolmogorov Flow and Navier-Stokes | prediction accuracy | Improved |
| reduce model zoo size below 50 (ablation) | DynaDiff · Cylinder Flow and Lambda-Omega | prediction error | Hurt |
| increase number of seen environments beyond about 20% of total (ablation) | DynaDiff · Kolmogorov Flow and Navier-Stokes | prediction error | No clear effect |
| randomly permute 10% of neurons in the model zoo (ablation) | DynaDiff weight graph · Cylinder Flow | performance | No clear effect |
| swap FNO expert for WNO or UNO (ablation) | DynaDiff · Cylinder Flow | prediction performance | No clear effect |
| train prompter with variable observation length 1 to L | DynaDiff prompter · observation length sensitivity | prediction quality | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DPOT — DPOT: Auto-Regressive Denoising Operator Transformer for Large-Scale PDE Pre-Training (ICML 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Sparse Diffusion Autoencoder for Test-time Adapting Prediction of Complex Systems (NeurIPS 2025 · similarity 0.86)
- Generating Physical Dynamics under Priors (ICLR 2025 · similarity 0.86)
- Environment Inference for Learning Generalizable Dynamical System (NeurIPS 2025 · similarity 0.86)
- One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling (NeurIPS 2025 · similarity 0.86)
- Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models (ICLR 2025 · similarity 0.85)
- Neural Context Flows for Meta-Learning of Dynamical Systems (ICLR 2025 · similarity 0.85)
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.
