Research map / Vision-language and multimodal / Multimodal LLMs
Qwen-3D: A Generalist 3D Vision-Language Model for Spatial Understanding
Lucy Lin, Ayush Jain, Yifan Liu, Katerina Fragkiadaki
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 |
|---|---|---|---|
| feed full text and visual tokens to mask decoder instead of a <REF> token (ablation) | Qwen-3D · ScanRefer Acc@50 (3B ablation) | Top-1 Acc: 39.7 -> 53.5 | Improved |
| use full attention mask instead of causal mask (ablation) | Qwen-3D · ScanRefer Acc@50 (3B ablation) | Top-1 Acc: 36.4 -> 53.5 | Improved |
| fine-tune the VLM backbone instead of freezing (ablation) | Qwen-3D · ScanRefer Acc@50 (3B ablation) | Top-1 Acc: 37.0 -> 53.5 | Improved |
| use aligned 3D RoPE instead of naive 3D MRoPE (ablation) | Qwen-3D · ScanRefer Acc@50 (3B ablation) | Top-1 Acc: 49.9 -> 53.5 | Improved |
| use aligned 3D RoPE instead of 2D RoPE (ablation) | Qwen-3D · ScanRefer Acc@50 (3B ablation) | Top-1 Acc: 53.2 -> 53.5 | No clear effect |
| joint 2D-3D training retains 2D capability | Qwen-3D 3B · RefCOCO / RealWorldQA | accuracy: RefCOCO 89.1 -> 88.1; RefCOCO+ 82.4 -> 82.7; RealWorldQA 62.6 -> 60.4 | Mixed |
| use Qwen-3D mask decoder instead of Grounded-3D-LLM | Qwen-3D 3B · ScanNet200 instance segmentation | mAP: 12.1 -> 25.3 | Improved |
| Qwen-3D vs specialist UniVLG out-of-domain | Qwen-3D 3B · Locate3D ScanNet++ | Acc@25: 32.3 -> 55.7 | Improved |
| Qwen-3D vs UniVLG on outdoor vehicle split | Qwen-3D 3B · 3EED Vehicle | A@0.25: 28.6 -> 22.9 | Hurt |
| use voxel pooling instead of token merging (ablation) | Qwen-3D · ScanNet dense scenes | memory | Improved |
| Qwen-3D 3B vs 7B | Qwen-3D · ScanRefer Acc@25 | accuracy: 62.2 -> 62.8 | No clear effect |
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.
- LLaVA-3D — LLaVA-3D: A Simple yet Effective Pathway to Empowering LMMs with 3D Capabilities (ICCV 2025) · better on 4 of 5 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
- Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering (NeurIPS 2026 · similarity 0.92)
- Video-3D LLM: Learning Position-Aware Video Representation for 3D Scene Understanding (CVPR 2025 · similarity 0.91)
- Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.91)
- Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning Synergy (ICLR 2026 · similarity 0.91)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.91)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.91)
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.
