Automated recognition of student engagement in online learning is crucial as it enables teachers to adapt content delivery to improve learning. In this paper, we explore a method that finetunes a pretrained vision language model (VLM) to recognize student engagement markers in still images. Our model learns to avoid incorrect answers during finetuning by using the emerging direct preference optimization techniques on self-generated preference pairs based on the correct and incorrect VLM answers. On publicly available student engagement datasets, our model shows superior performance over other approaches and substantially better generalizability over the traditional vision methods.