Predictive Emotional Selfhood in Artificial Minds (PESAM): A Unified and Synergistic Variational Framework

Abstract A grand challenge in artificial intelligence and neuroscience is to formally integrate emotion and selfhood into a unified, predictive model of the mind. Without such a framework, creating truly adaptive agents or understanding the computational basis of psychiatric disorders remains elusive. This paper introduces Predictive Emotional Selfhood in Artificial Minds (PESAM), a variational framework demonstrating that an emotional self emerges from the synergistic interaction of three core mechanisms: (1) Affective Precision Control (APC), the emotional modulation of sensory gain; (2) Self-as-Hyperprior (SaH), a deep, stabilizing self-model; and (3) Affective Homeostatic Objectives (AHO), intrinsic drives for internal stability. We argue that a true test of a unified framework lies in its ability to solve complex problems that are intractable for any single mechanism alone. To this end, we introduce a novel Social Threat & Body-Boundary Task. Results from this unified task show that the complete PESAM agent achieves significantly higher performance than both lesioned variants and strong alternative models (e.g., reinforcement learning), providing quantitative evidence for a genuinely synergistic—rather than merely additive—account of emotional selfhood and a principled foundation for adaptive AI and computational psychiatry.