Reinforcement learning models treat the basal ganglia (BG) as an actor–critic network. The ventral pallidum (VP) is a major component of the BG limbic system. However, its precise functional roles within the BG circuitry, particularly in comparison to the adjacent external segment of the globus pallidus (GPe), remain unexplored. We recorded the spiking activity of VP neurons, GPe cells (actor) and striatal cholinergic interneurons (critic) while monkeys performed a classical conditioning task. Here, we report that VP neurons can be classified into two distinct populations. The persistent population displayed sustained activation following visual cue presentation, was correlated with monkeys’ behavior and showed uncorrelated spiking activity. The transient population displayed phasic synchronized responses that were correlated with the rate of learning and the reinforcement learning model’s prediction error. Our results suggest that the VP is physiologically different from the GPe and identify the transient VP neurons as a BG critic.