Failure of Optimization-Based Multimodal Sentiment Analysis Balancing

Evaluations on CMU-MOSI and CMU-MOSEI datasets show that gradient and loss-based balancing strategies fail to outperform late concatenation in multimodal sentiment analysis. The research identifies that current methods conflate fitting speed with discriminative importance, leaving modality imbalance unresolved.