Multi-Modal Isotropic Perturbations for Cross-Attentive Latent Spaces

Existing regularization techniques are siloed by modality, failing to address the complex alignment issues in cross-attentive token spaces. We introduce a framework that injects coordinated isotropic stochastic noise across disparate text, vision, and audio embeddings simultaneously. By scaling the perturbation variance proportionally to the cross-modal attention weights. This cross-latent smoothing forces the model to bind concept representations geometrically rather than memorizing exact data streams, establishing new benchmarks for zero-shot multi-modal transfer tasks.