Fig. 1

Design and RT. a The first training session was proctored 3 days before the fMRI scan, while the second training session was proctored immediately before the scan. To reduce interference, fractal stimuli in each session were masked to be either squares or diamonds in shape. The fMRI scan included four runs for each of the two training sets. b For predictive actions, given the cue, one outcome appeared with 95% probability when the left button was pressed and the other with 95% probability when the right button was pressed. For nonpredictive actions, given the cue, either outcome appeared with 50% probability when the left or right button was pressed. c In verbal tests outside the scanner, participants spoke aloud either “top” or “bottom” to indicate which outcome was expected given the cue and action. d For each trial in the fMRI scanner, participants chose to press either a left-hand or right-hand button to replace a cue with an outcome. While choice RT was similar for predictive and nonpredictive actions immediately after training, it was faster for predictive vs. nonpredictive actions after a 3-day delay. Error bars indicate ±1 SEM of the difference between predictive and nonpredictive actions at each timescale. ***p < 0.001; *p < 0.05 (paired t-tests). Source data are provided as a Source Data file