微调 VLM 提升 AI 空间智能：理解 3D 与 2D 旋转
Fine-Tuning VLM for Enhancing AI's Spatial Intelligence: Understanding 3D and 2D Rotations
AI 导读
arXiv 一项研究通过多个物体旋转数据集微调视觉语言模型（VLM），在 2D 和 3D 旋转检测上取得明显改进。微调后的 Gemma-4 MoE 模型在按轴和角度预测旋转的任务上显著优于微调后的 Gemma-4 通用模型，且无需显式坐标系即可提升 2D 角度估计。实验还发现，具有明显线性特征的物体表现更好，而可识别物体并未提高角度检测精度。
正文 · 原文
Abstract:Spatial intelligence is a fundamental skill in multiple domains, such as Science, Technology, Engineering, and Mathematics (STEM), Medicine, Architecture, and Construction. Recent studies indicate that Vision-Language Models (VLMs) still face limitations in spatial reasoning, which inhibits artificial intelligence (AI) from performing practical spatial tasks. Using multiple object-rotation datasets developed for training and evaluation, our experiments demonstrated promising improvements in both 2D and 3D rotation detection. Fine-tuned Google DeepMind-built Gemma-4 mixture-of-experts (MoE) models significantly outperformed fine-tuned Gemma-4 generalist models in predicting rotations defined by both their axes and angles. Fine-tuning also substantially improved angle estimation for 2D representation without requiring an explicit coordinate system. Furthermore, identifiable objects did not improve angle-detection accuracy; instead, objects with prominent linear features showed improved performance.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.04206 [cs.AI] |
| (or arXiv:2610.04206v3 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.04206 arXiv-issued DOI via DataCite |
Submission history
From: Uttamasha Monjoree [view email]
[v1]
Sat, 3 Oct 2026 01:44:27 UTC (1,497 KB)
[v2]
Tue, 6 Oct 2026 02:46:30 UTC (936 KB)
[v3]
Fri, 30 Oct 2026 05:44:54 UTC (936 KB)
来源：arXiv cs.AI · arxiv.org