Macaque Area LIP Reflects Confidence-Dependent Changes in Decision Policy

The process of forming a decision gives rise to an estimate of its quality or likelihood of success. This graded sense of confidence is important for guiding subsequent decisions, but a neural mechanism linking subjective confidence to formation of the next decision has not been identified. We trained two male rhesus macaques to report a perceptual choice and simultaneous confidence judgment in a reaction-time (RT) motion discrimination task. Monkeys were more likely to repeat a rewarded choice when they reported low confidence on the previous trial, and showed greater changes in RT after a surprising outcome (low-confidence correct or high-confidence error). Neural activity in the lateral intraparietal area (LIP) encoded the previous trial's choice and confidence more strongly after a low-confidence correct trial, and the strength of history encoding was correlated with a trial-by-trial estimate of choice bias. Ramping dynamics of the decoded decision variable also depended systematically on the conjunction of confidence and reward on the previous trial, in a manner that reflected individual differences between animals. The findings suggest that LIP not only reflects the formation of the current decision but could participate in confidence-guided learning via adjustments of the subsequent decision process.Significance statement Decision confidence is of primary interest in the study of metacognition, yet it also plays a key role in learning, serving as a graded prediction of accuracy that modulates the response to feedback. Reanalyzing data from nonhuman primates performing a reaction-time (RT) motion discrimination task with explicit confidence reports, we found that adjustments of choice bias and RT were gated by the animal's self-reported confidence on the previous trial. The interaction of confidence and outcome (reward or punishment) was reflected in the activity of parietal (LIP) neurons on the subsequent trial, in ways largely consistent with the behavioral effects. The results and accompanying computational model suggest plausible functional links between the neuronal substrates of perceptual decision making and reinforcement learning.