Research map / Architectures and efficiency / RNNs and normalization
Walrus: A Cross-domain Foundation Model for Continuum Dynamics
Michael McCabe, Payel Mukhopadhyay, Tanya Marwah, Bruno Régaldo-Saint Blancard, François Rozet, Cristiana Diaconu, Lucas Thibaut Meyer, Kaze W. K. Wong, Mohammad-Hadi Sotoudeh, Alberto Bietti, Irina Espejo Morales, Rio Alexa Fear and 13 more
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 patch jittering during autoregressive rollout (ablation) | Walrus · pretraining datasets (19), long-horizon rollout | median VRMSE: 1.07 -> 0.34 (Maze) | Improved |
| add patch jittering (ablation) | Walrus · pretraining datasets (19) | unstable rollouts (VRMSE >= 1): 10 -> 3 of 19 datasets | Improved |
| add patch jittering (ablation) | Walrus · TRL (2D) full trajectory | median VRMSE: 76.9 -> 0.75 | Improved |
| add patch jittering (ablation) | Walrus · MultiQuadrantsP full trajectory | median VRMSE: 0.60 -> 0.62 | No clear effect |
| add patch jittering (ablation) | Walrus · Active Matter full trajectory | median VRMSE: 0.98 -> 0.98 | No clear effect |
| use topology-aware sampling plus adaptive tokenization and differential batch sizes (ablation) | Walrus training · distributed training throughput | throughput: +262% over naive FSDP | Improved |
| normalize token counts across datasets with adaptive-compute tokenization (ablation) | Walrus training · distributed training throughput | throughput | Hurt |
| use Walrus instead of prior foundation models (MPP, Poseidon, DPOT) | downstream finetuning · downstream tasks one-step | VRMSE: 63.6% average reduction | Improved |
| use Walrus instead of prior foundation models | downstream finetuning · downstream tasks medium-range | VRMSE: 48.3% average reduction | Improved |
| use Walrus instead of DPOT on linear acoustic tasks at long horizons | DPOT-H · Staircase avg T in [21:60] | VRMSE: 0.0031 vs 0.0074 | Hurt |
| add augmentation (2D to 3D embedding, time strides) in pretraining HalfWalrus (ablation) | HalfWalrus · pretraining metrics | pretraining loss | Hurt |
| add augmentation (2D to 3D embedding, time strides) in pretraining HalfWalrus (ablation) | HalfWalrus · unseen downstream tasks | VRMSE | Improved |
Closest papers by meaning
- The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning (NeurIPS 2024 · similarity 0.87)
- Space and time continuous physics simulation from partial observations (ICLR 2024 · similarity 0.87)
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (NeurIPS 2024 · similarity 0.86)
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.86)
- Poseidon: Efficient Foundation Models for PDEs (NeurIPS 2024 · similarity 0.86)
- Multiple Physics Pretraining for Spatiotemporal Surrogate Models (NeurIPS 2024 · 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.
