Laura De Laere^1,2, Ruth J. van Holst², Pascal Mamassian¹; ¹École Normale Supérieure Paris - PSL, ²Amsterdam UMC - University of Amsterdam

Metacognition, or the ability to monitor and regulate one’s cognitive processes, is often expressed through decision confidence. In perceptual decision-making, confidence reflects subjective certainty and can guide behavior by shaping decision boundaries. However, it is unclear whether confidence is monitored continuously during information gathering, and how sampling costs influence this process. We hypothesized that confidence would flexibly track task demands, decreasing under greater difficulty or cost, and that confidence would be more sensitive to new evidence when sampling was costly. To test these hypotheses, we adapted a visual information-gathering task in which participants (n=25) revealed as many hidden cards of a deck as they wished before reporting the dominant color. Using a 2x2 design, we manipulated decision difficulty by varying color proportions (60% or 75% dominant color) and sampling cost (1 or 2 points per card revealed). After each revealed card, participants had a 30% chance of having to provide a temporary decision and confidence rating, allowing us to track dynamic updates in confidence. Participants earned rewards for correct decisions, and a greater reward if fewer cards were revealed. As expected, participants showed lower accuracy in the difficult condition and sampled more when trials were difficult or sampling cost was low. Confidence decreased under high cost and for difficult trials. Accuracy had the largest impact on confidence, with incorrect decisions decreasing confidence substantially. A linear mixed-effects model predicting confidence based on evidence strength, cost, difficulty, and accuracy revealed that confidence increased with stronger evidence. Importantly, participants’ confidence ratings were more strongly influenced by new evidence when further sampling was discouraged by higher cost. Together, these results support our hypotheses and highlight that participants strategically calibrate their confidence. They reduced it when the task is harder or costlier, and weighted evidence more heavily when further sampling is discouraged.

Acknowledgements: MSCA Doctoral Network "CODE" (grant number: 101119647)