The paper argues that general spatial intelligence needs both static 3D and dynamic 4D understanding, but existing models tend to favor one at the expense of the other. The authors find this interference is non-uniform across parameter groups, with late-layer MLPs being especially sensitive to changes that shift the static–dynamic balance.
To address this, they propose DynThinker. Its Decoupled Exponential Moving Average (D-EMA) applies group-wise averaging to constrain drift in interference-sensitive pretrained parameters while still allowing flexibility for capability gains. A complementary Spatial-Dynamic Fusion (SDF) module selectively injects spatial and temporal priors from a dynamic geometry encoder into intermediate VLM representations under semantic guidance.
Across ten metrics from nine static 3D and dynamic 4D benchmarks, DynThinker achieves the highest average score among evaluated models. Notably, it records 56.0 on ReVSI, 67.2 on DSR-Bench, and 51.2 and 40.1 on the sample-wise and group-wise protocols of DSI-Bench, respectively, setting new state-of-the-art results among compared open-source models in a single unified framework.