Multimodal classifiers can converge to modality-dominant solutions in which one modality dominates the joint prediction, suppressing the learning of others. Existing balancing methods mainly adjust losses, gradients, or modality contributions, largely treating modality imbalance as an optimization problem while implici...
Ze-Chang Xiong, Da Li, Rong Yin et al.· 0 citations
STEPS is proposed, a selective OPSD framework that controls the location, direction, and duration of distillation that improves mean accuracy over GRPO and support brief, targeted self-distillation as an efficient complement to reward-driven reasoning training.
Jia-Xuan Wang, Xuan-Xin Ouyang, Z. Chen et al.· 15 citations
This work constructed Illusion-Reasoning, a benchmark of illusion images collected from the real world, incorporating diverse annotated question-answer pairs, and shows that the reasoning capabilities of a wide range of LVLMs are not as advanced as claimed.