Research map / Vision-language and multimodal / Multimodal LLMs
Mamba-3VL: Taming State Space Model for 3D Vision Language Learning
Yuan Wang, Yuxin Chen, Zhongang Qi, Lijun Liu, Jile Jiao, Xuetao Feng, Yujia Liang, Ying Shan, Zhipeng Zhang
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 IDPA to spatial+channel scanning (ablation) | Mamba-3VL · ScanRefer | accuracy (%): 52.4 -> 53.5 | Improved |
| add IDPA to spatial+channel scanning (ablation) | Mamba-3VL · SQA3D | accuracy (%): 51.0 -> 52.4 | Improved |
| add relation-prioritized spatial scanning to channel only (ablation) | Mamba-3VL · SQA3D | accuracy (%): 47.0 -> 51.0 | Improved |
| add relation-prioritized spatial scanning to channel only (ablation) | Mamba-3VL · Multi3DRefer | accuracy (%): 49.7 -> 52.0 | Improved |
| add channel twisting to spatial scanning (ablation) | Mamba-3VL · ScanRefer | accuracy (%): 50.1 -> 52.4 | Improved |
| add spatial scanning to IDPA-only (ablation) | Mamba-3VL · ScanRefer | accuracy (%): 49.5 -> 52.1 | Improved |
| use relation twisting instead of random-order scanning (ablation) | Mamba-3VL · Scan2Cap | CIDEr: 80.8 -> 84.0 | Improved |
| use relation twisting instead of random-order scanning (ablation) | Mamba-3VL · Nr3D Hard | accuracy (%): 58.3 -> 62.9 | Improved |
| add FIS component to scanning (ablation) | Mamba-3VL · Nr3D Hard | accuracy (%): 60.5 -> 62.9 | Improved |
| add NIS component to scanning (ablation) | Mamba-3VL · Nr3D Hard | accuracy (%): 61.7 -> 62.9 | Improved |
| use relation twisting instead of relation AdaIN (ablation) | Mamba-3VL · Scan2Cap | CIDEr: 83.1 -> 84.0 | Improved |
| replace Mamba-3VL instance-specific PE with linear PE (ablation) | Mamba-3VL · ScanRefer, Multi3DRefer, SQA3D, Scan2Cap, Nr3D | accuracy | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- 3D-VisTA — 3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment (ICCV 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Abstract 3D Perception for Spatial Intelligence in Vision-Language Models (CVPR 2026 · similarity 0.90)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.89)
- Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models (CVPR 2026 · similarity 0.89)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.89)
- HiSpatial: Taming Hierarchical 3D Spatial Understanding in Vision-Language Models (CVPR 2026 · similarity 0.89)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.89)
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.
