Emotions as Transferable Policy Primitives for General Intelligence

We propose a novel neural architecture that leverages discrete emotional tokens as transferable policy primitives for general intelligence. Building upon the theoretical framework thatemotions solve credit assignment in continuous action spaces, we present a practical implementation where emotional states serve as an interface between task-agnostic strategic reasoning andtask-specific behavioral execution. Our architecture separates learning into two components: (1)a universal emotion-selection policy trained via Monte Carlo Tree Search that transfers acrossdomains, and (2) task-specific reflex actors that map emotion-state pairs to concrete actions.This decomposition enables efficient transfer learning—strategic reasoning learned in one domain (e.g., Minecraft) directly applies to novel domains (e.g., FPS games) with only the reflexlayer requiring retraining. We formalize this architecture mathematically and demonstrate whythis separation is both computationally necessary and sufficient for achieving general intelligencein continuous control tasks