Spatial-Dynamic Interference
The phenomenon in which static 3D spatial understanding and dynamic 4D scene reasoning interfere with each other during model training, so that gains in one come at the expense of the other. The root cause is that the two capabilities share parameter pathways, with particular sensitivity in the later-layer MLPs.
Timeline (last 90 days)
There is a pronounced "spatial–dynamic interference" phenomenon between static 3D spatial understanding and dynamic 4D scene reasoning: post-training on 3D tasks weakens a model's ability to understand dynamic scenes, while models specialized in 4D reasoning often lack a solid foundation in static 3D spatial reasoning.
The balance between static and dynamic capabilities is particularly sensitive to changes in the parameters of the model's late-layer MLPs (multilayer perceptrons).
When a model is post-trained on 3D tasks, the high-level semantic pathways in late layers are preferentially reshaped, thereby overwriting the representations that originally supported dynamic temporal reasoning.
All Facts (3)
There is a pronounced "spatial–dynamic interference" phenomenon between static 3D spatial understanding and dynamic 4D scene reasoning: post-training on 3D tasks weakens a model's ability to understand dynamic scenes, while models specialized in 4D reasoning often lack a solid foundation in static 3D spatial reasoning.
50%UnverifiedThe balance between static and dynamic capabilities is particularly sensitive to changes in the parameters of the model's late-layer MLPs (multilayer perceptrons).
50%UnverifiedWhen a model is post-trained on 3D tasks, the high-level semantic pathways in late layers are preferentially reshaped, thereby overwriting the representations that originally supported dynamic temporal reasoning.
50%