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

A team of researchers has created an AI model that spontaneously exhibits behavior associated with the mirror self-recognition test, a classic benchmark for self-awareness. The model, developed by Dongmin Kim, Hoshinori Kanazawa, and Yasuo Kuniyoshi, relies on a core mechanism called a 'self-prior.' This self-prior is implemented using a Transformer neural network that learns the statistical patterns of an agent's own normal, multisensory experiences—essentially building a probabilistic model of 'self.' When a novel element, like a sticker on the face, appears in the mirror, it creates a discrepancy from this learned model. This mismatch drives the agent's behavior through a process called active inference, which seeks to minimize surprise or prediction error.

In simulations, a virtual agent equipped only with vision and proprioception (no tactile sense) successfully located and removed a sticker placed on its own face in approximately 70% of cases, all without any explicit training or external reward. The study confirmed that the act of removing the sticker significantly reduced the agent's 'expected free energy,' a formal measure of surprise, indicating the self-prior functioned as an internal criterion for distinguishing self from non-self. The model also demonstrated that the self-prior captures cross-modal associations, like linking visual appearance with body position, forming what the authors describe as a probabilistic body schema.

This work provides a concise, mechanistic account of a complex cognitive behavior using the principles of active inference and the free energy principle. It suggests that sophisticated, seemingly self-aware actions can emerge from a single, unified drive to maintain a coherent internal model of one's own body and sensory state. The code for the model is publicly available, inviting further research into the developmental origins of selfhood in artificial and biological systems.