generator-aligned-representation-interface-a8da335c·1 events·first seen Aliases: Generator-Aligned Representation Interface
Researchers introduce the Generator-Aligned Representation Interface (GARI), a design principle that exposes transformation generators to generic sequence backbones without requiring group-specific architectural changes. The framework formalizes soft equivariance via a probe-specific residual metric and instantiates it as GARI-Net, validated on genomic sequences, images, and 3D point clouds across symmetry transformations including reversal, rotation, and reflection. A companion diagnostic metric, Direct Equivariance Error (DEE), enables frozen-checkpoint evaluation of representation equivariance. The work offers a portable alternative to hard-equivariant architectures, making symmetry structure learnable and measurable across modalities.