Research map / Architectures and efficiency / Attention and transformers
Less Tokens, Better Forecasts: Sparse Residual Routing for Efficient Weather Prediction
Janet Wang, Yunbei Zhang, Lin Zhao, Xi Xiao, Jihun Hamm, Xiao Wang
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 Sparse-Reslim routing (r=0.25, split 2,8,2) | Res-Slim-ViT · ERA5 1.0° Z500 | RMSE: 868.7 -> 768.3 | Improved |
| add Sparse-Reslim routing | Res-Slim-ViT · ERA5 1.0° T2m | RMSE: 3.19 -> 2.68 | Improved |
| add Sparse-Reslim routing | Res-Slim-ViT · ERA5 0.25° Z500 | RMSE: 1042 -> 892 | Improved |
| add Sparse-Reslim routing to EDM generative model | EDM ViT · ERA5 1.0° Z500 | RMSE: 951.3 -> 889.7 | Improved |
| add Sparse-Reslim routing to EDM generative model | EDM ViT · ERA5 1.0° | training throughput (samples/s/GPU): 11.4 -> 20.8 | Improved |
| add Sparse-Reslim routing | Res-Slim-ViT · ERA5 1.40625° T2m 10-day rollout | ACC: 0.53 -> 0.69 | Improved |
| lower keep ratio from 0.25 to 0.10 (ablation) | Sparse-Reslim · ERA5 1.0° Z500 | RMSE: 768.3 -> 801.5 | Hurt |
| raise keep ratio to 1.0 (dense) (ablation) | Sparse-Reslim · ERA5 1.0° Z500 | RMSE: 768.3 -> 868.7 | Hurt |
| use one dense early and late block (1,10,1) instead of (2,8,2) (ablation) | Sparse-Reslim · ERA5 1.0° Z500 | RMSE: 768.3 -> 812.4 | Hurt |
| use four dense blocks per end (4,4,4) instead of (2,8,2) (ablation) | Sparse-Reslim · ERA5 1.0° Z500 | RMSE: 768.3 -> 831.2 | Hurt |
| replace random selection with top-K attention score (ablation) | Sparse-Reslim · ERA5 1.0° Z500 | RMSE: 768.3 -> 784.1 | Hurt |
| replace random selection with learned difficulty head (ablation) | Sparse-Reslim · ERA5 1.0° Z500 | RMSE: 768.3 -> 779.8 | Hurt |
Closest papers by meaning
- Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting (NeurIPS 2026 · similarity 0.85)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.85)
- Scaling transformer neural networks for skillful and reliable medium-range weather forecasting (NeurIPS 2024 · similarity 0.85)
- DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.84)
- CirT: Global Subseasonal-to-Seasonal Forecasting with Geometry-inspired Transformer (ICLR 2025 · similarity 0.84)
- Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution (CVPR 2025 · similarity 0.84)
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.
