Sijie Mai

A unified model that adapts to the interplay between perceptual fidelity and prompt alignment can achieve state-of-the-art performance while revealing interpretable insights into human judgment.

Text modality dominance is holding back multimodal models, but penalizing gradient conflicts and aligning predicted statistics can unlock synergistic training and state-of-the-art results.

Multimodal models can now shrug off noisy or shifted data thanks to a new method that disentangles causal, stable signals from spurious correlations.

By unifying noisy and missing modalities into a single "low-quality" problem, UMQ provides a surprisingly effective framework for robust multimodal affective computing.