Research map / Architectures and efficiency / Attention and transformers
MeshTok: Efficient Multi-Scale Tokenization for Scalable PDE Transformers
Yanshun Zhao, Xiaoyu Peng, Jiamin Jiang, Congcong Zhu, Jingrun Chen
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 activity-based adaptive refinement (top 25% patches) to coarse tokenization (ablation) | MeshTok Ours-Big · Averaged benchmarks, 1-step | relative L2 error: 1.138 -> 0.972 | Improved |
| add activity-based refinement (ablation) | MeshTok Ours-Big · Averaged benchmarks, 10-step rollout | relative L2 error: 2.607 -> 2.261 | Improved |
| refine all patches (full refinement) instead of partial refinement (ablation) | MeshTok Ours-Small · Averaged benchmarks, 1-step | relative L2 error: 1.979 -> 2.014 | Hurt |
| refine all patches instead of partial refinement (ablation) | MeshTok Ours-Large · Averaged benchmarks, 1-step | relative L2 error: 0.837 -> 0.716 | Improved |
| match MACs by widening or deepening a coarse-token baseline (ablation) | Uniform coarse ViT-style baseline · Compute-matched, Ours-Big | relative error: 0.9723 (MeshTok) vs 1.0294 and 1.0652 | Hurt |
| full refinement with reduced width or depth to match MACs (ablation) | Uniform full-refinement baseline · Compute-matched, Ours-Big | relative error: 0.9723 (MeshTok) vs 1.0777 and 1.0632 | Hurt |
| pretrain on source PDEs before fine-tuning (ablation) | MeshTok · Shear flow, Burgers, Reaction-Diffusion | relative L2 error: Shear 3.817 -> 3.334; Burgers 0.324 -> 0.252; React-Diff 11.096 -> 8.106 | Improved |
| pretrain on source PDEs before fine-tuning (ablation) | MeshTok · Black-Scholes-Barenblatt | relative L2 error: 0.127 -> 0.142 | Hurt |
| increase model scale from Small to Large with AMR refinement (ablation) | MeshTok · Averaged benchmarks, 1-step | relative L2 error: 1.639 -> 0.837 | Improved |
| use MeshTok instead of BCAT | PDE Transformer · PDEBench CNS (0.1,0.01) | relative L2 error: 0.254 -> 0.249 | No clear effect |
| use MeshTok instead of BCAT | PDE Transformer · PDEBench SWE | relative L2 error: 0.443 -> 0.276 | Improved |
| add AMR refinement in 3D (k = 0.375) versus no refinement (ablation) | MeshTok · The Well MHD (3D) | relative error: 39.647 -> 35.851 | Improved |
Closest papers by meaning
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations (ICML 2025 · similarity 0.88)
- TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters (ICLR 2025 · similarity 0.86)
- Efficient Generative Transformer Operators for Million-Point PDEs (NeurIPS 2026 · similarity 0.86)
- Scaling Mesh Generation via Compressive Tokenization (CVPR 2025 · similarity 0.85)
- Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries (ICML 2025 · similarity 0.85)
- Transolver: A Fast Transformer Solver for PDEs on General Geometries (ICML 2024 · 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.
