Conference on Cognitive Computational Neuroscience

Transient goals are key motivators in human learning. Extending the classic reinforcement learning framework to include a flexible mapping of outcomes to rewards according to current goals accounts for goals as intrinsic reinforcers. However, learning by encoding transient goals as rewards is slower than learning with familiar rewards. Here, we test whether this effect is due to occasional lapses in goal encoding and the subsequent value updating. We tasked human participants with learning from both familiar rewards (the "Points" condition) and abstract novel "goal" images (the "Goals" condition). To detect lapses in goal encoding, we asked participants to report all positive outcomes they received. Behavioral results replicated our previous finding that people learn less efficiently when they encode goals as rewards than directly receiving familiar rewards. However, our modeling analysis suggested that lapses could not fully explain this behavioral discrepancy. This finding challenges the hypothesis that lapses are the primary cause of slower goal-driven learning, providing insights into the complex cognitive mechanisms of how humans learn from flexible, goal-dependent value assignments.