Deinforcement Learning: Rethinking Reward-Based Learning Mechanisms

Conventional reinforcement learning (RL) paradigms, rooted in dopaminergic reward prediction, conflate the absence of reward with punishment. This simplification overlooks pain, a biologically distinct aversive signal with dedicated neural pathways and learning mechanisms. Here we introduce Deinforcement Learning, a framework that integrates pain as an independent computational driver of behavior. Drawing on evidence from neuroscience, psychology, and cognitive science, we show how pain circuits provide rapid, high-salience signals for avoidance and safety, complementing reward-based adaptation. We formalize this perspective in the MaxPain model, which rebalances learning between appetitive and aversive drives. This reconceptualization advances biologically grounded models of decision-making and suggests algorithms with faster convergence, improved robustness, and safer exploration.