As aggressive driving behavior threatens road safety, we investigate it using attention-based models and feature extraction techniques, applied on the METEOR driving dataset. After refining the dataset for our specific research needs, we implemented and evaluated the attention-based models OadTR and Colar, each showing distinct strengths and limitations. Notably, there’s a consistent correlation between the frequency of an action in the training data and a model’s classification accuracy. Our research offers two primary contributions. Firstly, we combined the OadTR and Colar models into a novel hybrid architecture that leverages categorical exemplars while predicting future frames. Secondly, we extract salient cues for model interpretability by tracing agent paths across spatial and temporal dimensions. These insights are especially valuable for autonomous vehicle applications where real-time interpretability and efficient computation are vital. The code is made available at https://github.com/unofficial-Jona/assessing_ADB/tree/main .