Abstract
Confidence judgment on one visual task can influence the confidence on the following task, a phenomenon known as “confidence leak” (Rahnev et al., 2015). However, little is known about the metacognitive mechanisms underlying this serial dependence in confidence judgments across multiple trials. In the present study, we investigated the mechanisms underlying the serial dependence of visual confidence using random-dot motion patterns as stimuli. On each trial, observers discriminated the motion direction (left or right) and rated their confidence in their direction-discrimination response on a 4-point scale. Task difficulty was controlled by varying motion coherence and was calibrated to each observer’s direction discriminability. Trials with medium difficulty level (target trials) were always preceded by either one or two consecutive trials that were either easy or hard (i.e., hard-medium, hard-hard-medium, easy-medium, easy-easy-medium, etc.). In Experiment 1, we found that confidence ratings for target trials (perceptual performance matched) were higher when these trials were immediately preceded by easy trials than by hard trials. In Experiment 2, this serial-dependence effect disappeared when observers rated the motion speed (a non-confidence response) in the immediately-preceding trial. In Experiment 3, the serial-dependence effect remained minimal even when the motion speed in the preceding trial was manipulated to enhance motor priming effect. We compared candidate models that updated decision criteria through different processes. Model evidence was higher 1) when decision criteria were updated based on a template set of criteria on every trial (template-updating models) than when criteria were updated based on those used in recent trials (serial-updating models), and 2) when considering only explicit confidence ratings from preceding trials and ignoring implicit confidence estimates in non-confidence, speed-rating trials than when considering both. Our findings suggest that the metacognitive system is adaptive in response to recent task and response history.