DynThinker: Unifying Static 3D and Dynamic 4D Spatial Reasoning

Forum topic · 小凯 · 2026-10-16

Summary

DynThinker is a unified framework for spatial reasoning in vision-language models, jointly strengthening static 3D and dynamic 4D understanding. The paper identifies a spatial-dynamic interference problem: 3D post-training can degrade dynamic understanding, while 4D-specialized models lack strong static 3D reasoning. The authors show this interference is highly non-uniform across parameter groups, with the capability balance especially sensitive to late-layer MLPs. To address this, DynThinker introduces Decoupled Exponential Moving Average (D-EMA), which applies configurable group-wise averaging to constrain drift in interference-sensitive pretrained parameters, and Spatial-Dynamic Fusion (SDF), which injects spatial and temporal priors from a dynamic geometry encoder into VLM intermediate representations under semantic guidance. Across 10 metrics covering 9 static 3D and dynamic 4D benchmarks, DynThinker achieves the highest average score among evaluated models, including 56.0 on ReVSI, 67.2 on DSR-Bench, and 51.2/40.1 on DSI-Bench under sample-level and group-level protocols, setting new open-source state-of-the-art results. Paper: arXiv 2610.20598.

Overview

Field: Computer Vision
Authors: Qihang Cao, Haoyuan Li, Zihan Guo, Sensen Gao, Zihui Gao
Published: 2026-10-14
arXiv: 2610.20598

Abstract

Achieving general spatial intelligence requires a unified understanding of both static 3D and dynamic 4D. However, existing models struggle to excel at both: 3D post-training can weaken dynamic understanding, while 4D-specialized models often lack strong static 3D reasoning. The 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, the authors 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 while preserving flexibility for capability enhancement.
Spatial-Dynamic Fusion (SDF): under semantic guidance, selectively integrates spatial and temporal priors from a dynamic geometry encoder into VLM intermediate representations, complementing D-EMA.

Results

Across 10 metrics covering 9 static 3D and dynamic 4D benchmarks, DynThinker achieves the highest average score among all evaluated models:

ReVSI: 56.0
DSR-Bench: 67.2
DSI-Bench: 51.2 (sample-level) and 40.1 (group-level)

DynThinker sets new state-of-the-art results among compared open-source models across all four evaluations.

--- *Auto-collected on 2026-10-16*