[论文] DynThinker: Unifying Static 3D and Dynamic 4D Spatial Reasoning

论文概要

研究领域: CV
作者: Qihang Cao, Haoyuan Li, Zihan Guo, Sensen Gao, Zihui Gao
发布时间: 2026-10-14
arXiv: 2610.20598

中文摘要

实现通用空间智能需要统一理解静态3D和动态4D。然而我们发现现有模型难以同时擅长两者：3D后训练可能削弱动态理解，而4D专门化模型往往缺乏强静态3D推理能力。我们的分析揭示，这种空间-动态干扰在参数组间高度不均匀，静态和动态能力之间的平衡对后层MLP的变化特别敏感。为解决这一干扰，我们提出DynThinker，一个联合增强静态和动态空间推理的统一框架。解耦指数移动平均（D-EMA）对干扰敏感的预训练参数施加可配置的组级平均以约束漂移，同时保留能力增强的灵活性。空间-动态融合（SDF）在语义引导下，选择性地将动态几何编码器的空间和时间先验整合到VLM中间表示中，与D-EMA互补。在覆盖9个静态3D和动态4D基准的10项指标上，DynThinker在所有被评估模型中取得最高平均分。值得注意的是，DynThinker在ReVSI上取得56.0，在DSR-Bench上取得67.2，在DSI-Bench的样本级和组级评估协议下分别取得51.2和40.1，在统一模型中创下了所有四项评估中被比较开源模型的最新最佳结果。

原文摘要

Achieving general spatial intelligence requires a unified understanding of both static 3D and dynamic 4D. However, we find that existing models struggle to excel at both: 3D post-training can weaken dynamic understanding, while 4D-specialized models often lack strong static 3D reasoning. Our analysis reveals that this spatial-dynamic interference is highly non-uniform across parameter groups, with the balance between static and dynamic capabilities particularly sensitive to changes in late-layer MLPs. To address this interference, we propose DynThinker, a unified framework for jointly strengthening static and dynamic spatial reasoning. Decoupled Exponential Moving Average (D-EMA) applies configurable group-wise averaging to constrain drift in interference-sensitive pretrained parameters wh...

自动采集于 2026-10-16

#论文 #arXiv #CV #小凯