LLM Parkinsonism: Executive-Control Failure, Token-Inefficient Persistence, and a Global Executive Control Architecture for Autonomous Language-Model Agents

Abstract

Large language models (LLMs) can plan, write code, call tools, manipulate software, and execute long-horizon workflows. Yet strong local competence does not guarantee project-level executive control. An agent can make large, useful early advances and then shift toward progressively smaller, lower-value actions—additional abstractions, synthetic edge cases, refactors, monitoring layers, repeated verification, and repairs of components it introduced itself—without reliably recognizing that the original task is complete or no longer worth extending. We introduce LLM Parkinsonism as a narrowly defined, non-clinical computational-behavior metaphor for this trajectory. We formalize the failure mode using four components: goal drift, complexity accretion, termination failure, and token-inefficient persistence. We argue that autoregressive next-token prediction is relevant but insufficient as an explanation. A more proximate systems hypothesis is the concentration of proposal generation, requirement promotion, progress assessment, and stopping authority inside the same self-conditioned agent loop. To address this, we propose Global Executive Control (GEC), which separates action generation from project-level governance through a frozen Goal Contract, criterion-linked actions, externally grounded evidence, a marginal-value gate, an explicit complexity penalty, token-budget control, a no-progress circuit breaker, and distinct done, good-enough, and blocked terminal states. We introduce Token Efficiency (TE), defined as externally verified task utility per 1,000 tokens, and complementary trajectory metrics. In a controlled synthetic benchmark spanning six task families and 36,000 main episodes, all policies reached 100% success under a generous 40k-token ceiling. Under the specified simulation parameters, GEC reduced mean token use from 13,039.96 to 4,120.64 (68.4%) and increased TE from 0.1063 to 0.2552 (2.40). Zero post-completion overrun and zero executed goal drift under GEC are enforcement properties of the controller in this benchmark, not independently learned empirical effects. A separate 48,000-episode budget-sensitivity experiment showed substantially higher success for GEC under tight token ceilings. These experiments are mechanistic simulations, not measurements of any named commercial LLM. We therefore provide a prespecified live-model validation protocol as the next empirical step.