Necessity of a 3D spatial latent representation for mental rotation

Determine whether rotation actions in the proposed mechanistic model of mental rotation must operate on the 3D spatial latent representation produced by the Equivariant Neural Renderer, or whether applying rotation actions directly to the neuro-symbolic sequence representation generated by the Vision Symbolic Model suffices to achieve accurate similarity judgments; specifically, ascertain if the 3D latent space is necessary for solving the Shepard–Metzler mental rotation task within this architecture.

Background

The paper introduces a three-module model for human mental rotation: an Equivariant Neural Renderer (EqNR) that yields a manipulable 3D spatial latent, a Vision Symbolic Model (VSM) that maps the spatial latent to a symbolic sequence under the Quadrant Hypothesis, and a decision MLP that selects rotation actions or similarity judgments based on pairs of symbolic descriptions.

While the authors argue for a hybrid spatial–symbolic account that matches behavioral signatures from their VR experiments, they note that in principle rotation actions could be implemented directly in symbolic space without using the 3D spatial latent. This raises the question of whether the spatial latent is strictly required for the task or if a purely symbolic pathway could suffice, which has implications for both cognitive theory and model design.