Prediction Errors Improve Declarative Learning but Not the Salience of Individual Words

Recent studies using the variable choice paradigm for a word-association learning task suggest that reward prediction errors (RPEs) enhance one-shot declarative learning. However, these studies may confound true associative learning with an “RPE bias,” where RPE items become more salient or fluent and are therefore chosen during recognition. This study tested whether such a bias exists and whether RPEs genuinely strengthen word-pair learning. Across three sessions, participants (n = 48) completed a modified version of the variable choice paradigm in which English-Swahili word pairs were learned under either RPE or no-RPE conditions. Results provided moderate evidence against an RPE bias: RPE words were not more likely to be selected when incorrect. In contrast, correctness strongly predicted choice, and correct translations associated with RPEs were moderately more likely to be chosen than correct translations without RPEs, indicating a true learning benefit. Confidence analyses mirrored this pattern, showing strong evidence against an effect of RPE bias but extreme evidence for an effect of correctness. Furthermore, confidence was much higher for correct translations with RPE compared to correct translations without RPE. Together, these findings demonstrate that RPEs improve associative memory without producing choice biases. Our results validate previous conclusions about RPE-driven declarative learning benefits.