VLM^3 shows that most of the key designs in 3D vision models turn out to be not important, these even include the regression formulation, where we can simply use discrete text tokens + cross entropy…