Token-Level Representation Learning for Robust Multimodal Model Generation. JDIAS 2026, 1 (3).