Multimodal Fusion: Reliability Scores Often Don't Influence Decisions

Many multimodal AI systems are designed to assess the reliability of individual data modalities and then weight their contributions to a final prediction accordingly. However, it has been unclear whether these reliability scores truly impact the model's decision-making process or merely reflect overall performance. Researchers have developed a straightforward diagnostic method to investigate this. After a model is trained and its inputs are fixed, the reliability scores are randomly permuted across test examples. If the model's predictions genuinely depend on these scores, its performance should degrade significantly. Experiments conducted on stress recognition (StressID) and sentiment analysis (CMU-MOSEI) datasets showed that permuting reliability scores often had no effect on performance. This suggests that, despite the potential benefits of selecting the best modality per example, the models' fusion rules were not effectively utilizing this reliability information. In contrast, positive control experiments, where reliability signals accurately identified the correct modality, demonstrated that the same fusion rules could yield substantial improvements, indicating that reliability signals only influence decisions when they are highly predictive of unimodal correctness.