Research map / Detection and segmentation / CNN features
MambaTree: Tree Topology is All You Need in State Space Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MambaTree, a state space model that propagates features over an input-dependent minimum spanning tree with a linear-time dynamic programming scan, applied to vision backbones (MambaTreeV) and LoRA-tuned LLMs (MambaTreeL).
Problem
Sequence-based SSMs (Mamba and its visual variants) rely on fixed scan orders that break spatial continuity and struggle with long-range dependencies in both images and text.
Approach
Build an input-aware minimum spanning tree over pixels or tokens (edge weights from feature cosine dissimilarity) and run Mamba-style state propagation along tree paths, with a two-pass dynamic programming algorithm that reduces the all-roots aggregation from O(L^2) to O(L).
Key results
- MambaTreeV-T reaches 83.4% top-1 on ImageNet-1K, 2.9 points above ViM-S and 0.9 above VMamba-T at similar FLOPs.
- MambaTreeV-T achieves 48.5 single-scale mIoU on ADE20K, 3.6 above ViM-S.
- MambaTreeL with LoRA-style tuning on Mamba-130M raises average zero-shot accuracy across 7 benchmarks from 46.1 (LoRA) to 47.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace raster scanning with tree topology scanning (ablation) | MambaTreeV-T · ImageNet-1K | accuracy: 82.6 -> 83.4 | Helped |
| Replace cross scanning with tree topology scanning (ablation) | MambaTreeV-T · ImageNet-1K | accuracy: 83.1 -> 83.4 | Helped |
| Use cosine rather than Manhattan distance for edge weights (ablation) | MambaTreeV-T · ImageNet-1K | accuracy: 82.9 -> 83.4 | Helped |
| Use cosine rather than Euclidean distance for edge weights (ablation) | MambaTreeV-T · ImageNet-1K | accuracy: 83.2 -> 83.4 | Helped |
| Traverse all vertices as roots rather than only the first vertex (ablation) | MambaTreeV-T · ImageNet-1K | accuracy: 82.9 -> 83.4 | Helped |
| Traverse all vertices as roots rather than only the last vertex (ablation) | MambaTreeV-T · ImageNet-1K | accuracy: 83.0 -> 83.4 | Helped |
| Share the same tree topology structure across stages (ablation) | MambaTreeV-T · ImageNet-1K | Top-1 accuracy: 83.4 -> 83.4 | No clear effect |
| Fine-tune with MambaTreeL rather than LoRA | Mamba 130M · Language benchmarks average | Average Acc.: +1.1% | Helped |
| Fine-tune with MambaTreeL rather than LoRA | Mamba 130M · WinoGrande | accuracy: 52.2 -> 51.1 | Hurt |
Surprising findings
- Restricting the tree-root traversal to only the first or last vertex lowers accuracy by only 0.4-0.5 points, so the all-roots traversal gives a modest gain.
Limitations
- The tree structure is not a common paradigm and must be optimized for specific hardware.
- Throughput is lower than VMamba-T and PlainMamba-L2 without a custom CUDA kernel.
- Language experiments are small in scale (Mamba-130M, Alpaca) and the authors note a marginal drop on WinoGrande.
Experimental setup
Datasets: ImageNet-1K MSCOCO 2017 ADE20K Alpaca
Benchmarks: ImageNet-1K MSCOCO 2017 ADE20K PIQA AI2-ARC SST WinoGrande LAMBADA RACE OpenBookQA
Metrics: Top-1 accuracy box mAP mask mAP mIoU accuracy throughput
Baselines: ViM VMamba LocalVMamba PlainMamba Swin Transformer ConvNeXt InternImage UniRepLKNet LoRA on Mamba-130M
Builds on: Mamba VMamba LocalMamba Vision Tree Scanning (MST-based feature propagation) LoRA
Closest papers by meaning
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.86)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.85)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.85)
- QuadMamba: Learning Quadtree-based Selective Scan for Visual State Space Model (NeurIPS 2024 · similarity 0.85)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.85)
- Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data (NeurIPS 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.