Online interviews have gained prominence in recruitment processes due to their time-saving nature and broader accessibility. To expand interview opportunities for candidates, we propose an Adaptive Multimodal Transformer for automated analysis. We introduce a two-tier evaluation scheme developed by domain experts to assess candidate performance in a structured and consistent manner. The task presents two major challenges: (1) missing acoustic input in part of the data, and (2) an imbalanced score distribution with most scores concentrated in the middle range. The proposed system evaluates candidates based on both textual and acoustic features. To address instances where acoustic input is unavailable, we use modality-specific control signals for missing inputs. Additionally, we propose a novel rebalancing optimization schedule to adjust training sample weights dynamically. Experimental results show that our approach predicts evaluation scores more accurately than other multimodal neural networks, achieving a mean squared error (MSE) of 0.476 and a Pearson correlation coefficient of 0.686.