Anisotropic Modality Align

Researchers introduced AnisoAlign, a geometric correction method for unpaired multimodal alignment.

The paper argues that modality gaps come largely from anisotropic residual directions, not just a simple shift.

AnisoAlign adjusts source representations toward the target modality while preserving semantics.

This reframes modality mismatch as a structured geometry problem that can be fixed in shared representation space.

The approach could reduce reliance on large paired datasets for training multimodal large language models.