Research map / Architectures and efficiency / Attention and transformers
Mamba-ND: Selective State Space Modeling for Multi-Dimensional Data
Shufan Li, Aditya Grover, Harkanwar Singh
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 |
|---|---|---|---|
| alternate scan orderings across layers (H+H-W+W-T+T-) instead of flattening into one 1D sequence (ablation) | Mamba-ND · HMDB-51 | top-1 accuracy: 34.9 -> 59.0 | Improved |
| alternate scan orderings across layers instead of 1D flattening (ablation) | Mamba-ND · ImageNet-1K val | top-1 accuracy: 76.4 -> 79.4 | Improved |
| alternating-directional design versus layer-level Multi-Head SSM / ND-SSM / Bi-SSM (ablation) | Mamba-ND · ImageNet-1K and HMDB-51 | top-1 accuracy: 79.4 vs 77.6, 77.2, 74.6 on ImageNet | Improved |
| layer-level Bi-SSM instead of single-direction 1D-SSM (ablation) | Mamba-ND · ImageNet-1K val | top-1 accuracy: 76.4 -> 74.6 | Hurt |
| group directions into parallel multi-directional blocks at block level (ablation) | Mamba-ND · HMDB-51 | top-1 accuracy: 59.0 -> 38.3 (hex-directional block) | Hurt |
| factorize the 3D scan into smaller sequences (ablation) | Mamba-ND · HMDB-51 | top-1 accuracy: 59.0 -> 55.8 (2D+1D) | Hurt |
| smaller patch size 8 instead of 16 (ablation) | Mamba-2D-S · ImageNet-1K val | top-1 accuracy: 79.4 -> 81.7 | Improved |
| set temporal SSM initial delta scale to 1.0 versus 5.0 (ablation) | Mamba-3D · HMDB-51 | top-1 accuracy: 55.3 -> 59.0 | Improved |
| change ordering of first three layers or swap TWH+ to WTH+ (ablation) | Mamba-3D · BTCV | DICE | No clear effect |
| Mamba-ND instead of Transformer baselines | Mamba-3D · ERA5 weather forecasting | ACC: 89.3 -> 90.1 | Improved |
| Mamba-3D versus Swin-UNETR on BTCV with a small model | Mamba-3D-UNETR-S · BTCV 3D segmentation | DICE: 83.3 -> 83.1 | No clear effect |
| add alternating directions to Bi-SSM (ablation) | Bi-SSM · ImageNet-1K and HMDB-51 | top-1 accuracy: +1.5 ImageNet, +8.4 HMDB-51 | Improved |
Closest papers by meaning
- Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data (NeurIPS 2025 · similarity 0.90)
- Block-Biased Mamba for Long-Range Sequence Processing (NeurIPS 2025 · similarity 0.89)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.88)
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.88)
- MambaTree: Tree Topology is All You Need in State Space Model (NeurIPS 2024 · similarity 0.88)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.88)
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.
