Active Inference with a Self-Prior in the Mirror-Mark Task

This study presents a computational model that utilizes a self-prior implemented with a Transformer to explain behavior in the mirror self-recognition test, demonstrating that a simulated infant can identify and remove a sticker from its face with approximately 70% success without external rewards. The model operates through active inference, where the discrepancy between learned multisensory experiences and novel stimuli drives mark-directed behavior, indicating a significant decrease in expected free energy post-removal. This research offers insights into the emergence of self-awareness and suggests that the free energy principle may unify theories on its developmental origins, making it relevant for practitioners exploring the intersection of AI and cognitive development.