Large Organismic Model: A Developmental Architecture for Embodied, Affective, and Self-Directed Artificial Intelligence

Abstract

We propose the Large Organismic Model (LOM), a developmental architecture for artificial intelligence. Rather than reporting the performance of an implemented system, this position paper formulates a testable computational hypothesis that integrates ideas from biology, developmental neuroscience, reinforcement learning, world models, developmental robotics, intrinsic motivation, and episodic memory. In contrast to the language-first developmental trajectory common in contemporary foundation models, LOM adopts a body-first developmental architecture: a persistent virtual body, interoceptive state, sensorimotor control, a body-conditioned world model, body-derived affective appraisal, episodic memory, and models of other agents are developed before language is introduced. Unless otherwise noted, "body" refers to a virtual body with persistent internal states rather than a biological body. LOM introduces a signed viability margin that quantifies the body's safety margin relative to the boundary of a viable region, and derives primitive valence from predicted changes in that margin. Nociception, fatigue, and resource deficits are not treated merely as externally assigned labels; instead, they are transformed into affective control signals through predictions of bodily and environmental consequences. Experiences are stored as episodes containing predictions, realized outcomes, affective evaluations, prediction errors, and provenance. Learning occurs on two time scales: rapid online adaptation and sleep-like offline consolidation involving replay, credit assignment, and counterfactual rollout. Because self-maintenance alone does not adequately account for spontaneous learning, LOM further defines a policy-relative viability set: the set of states the system can actually keep viable under its current skills and policy. Safe expansion of this set, together with empowerment, information gain, and learning progress, contributes to intrinsic motivation. The "free-will-like behavior" discussed here is not intended as a claim about philosophical free will. We use the operational term agentic autonomy to denote the capacity to generate and select exploratory goals from internal state, memory, and world-model predictions even in the absence of an immediate externally assigned task. We do not assume that experience of self-injury is sufficient to produce moral behavior. A self-generated pain-like aversive state may provide a computational basis for predicting harm to others through a body model, but avoiding harm to others additionally requires social appraisal, mutual viability, hard harm constraints, rights and fairness considerations, and related mechanisms. In the latter half of the paper, we discuss provenance-aware transmission of experience among multiple LOMs, cultural learning, and intergenerational inheritance. Finally, we interpret the J-space reported by Gurnee et al. (2026) through the lens of LOM and propose O-space (Organismic Workspace) as a hypothetical shared workspace spanning nine functional modules. We further advance the falsifiable hypothesis that affective salience acts as a gate regulating access to O-space.