ARCHIVES

In recent years there has been an increase in use of multimodal artificial intelligence (AI) systems, i.e., AI systems processing heterogeneous inputs simultaneously; they have proven effective empirically in many real-world applications, especially those related to safety critical tasks like disaster classification. However, there is a structurally underexplored failure mode in these systems which is the position bias. This paper reviews the existing research done on position bias in multimodal systems for disaster triaging and identifying the research gap that motivates the companion experimental paper. A detailed search was performed covering peer-reviewed publications from 2018 to 2025 and approximate total of 55 articles were reviewed completely based on their relevance to the four thematic areas (multimodal learning for disaster response, position and order bias in neural networks, consistency regularization, and fairness evaluation metrics). According to the results, while each thematic area has been independently studied, none of them specifically addresses the issue of positionally unstable multimodal disaster AI systems. Research into disaster triage methods commonly do not evaluate the fairness or stability of triage models. The literature on fairness in machine learning defines bias solely as demographic bias and does not take into consideration the sensitivity to modality contributions. Variance-based consistency regularization loss functions have been developed primarily for semi-supervised and contrastive learning paradigms and have yet to be implemented effectively in multimodal fusion paradigms where the ordering of input modalities vary during inference time.