Does signed prediction error drive declarative memory? Evidence from variable choice paradigms

Prediction error (PE) is the discrepancy between predictions and new information. For a binary reward outcome, PE may be signed (positive if the outcome was better than predicted and negative if the outcome was worse than predicted) or unsigned (absolute value of "surprise"). Using a "variable choice" paradigm, De Loof et al. (PLOS ONE, 131, Article e0189212, 2018) examined the role of PE in one-shot learning of unknown translations of known words and showed that associative memory for the translation was greater when (financial) reward was more unexpected and lesser when an expected reward was not received (i.e., signed PE); an effect that they replicated in several subsequent studies. However, other work on PE in declarative memory has assumed that memory is greater when an outcome is more unexpected, without any explicit reward (i.e., unsigned PE). We replicated De Loof et al.'s paradigm with and without financial reward, and found that memory was explained slightly better by unsigned PE (Experiments 1A-1B). However, we also identified a potential confound in the paradigm that could explain the results without any role of PE, as confirmed by simulations. We therefore designed a modified version of the paradigm that circumvents this confound (Experiment 2). Results were inconsistent with the PE account. We conclude that variable choice paradigms may not be well-suited to investigate the role of PE in one-shot declarative learning, and that the purported role of signed PE in declarative memory requires further investigation.