Fig. 2. Two-step task illustration.

A Participants navigated between the two stages of the task in order to reap rewards. The second stage included two pairs of fractal images, which led probabilistically to a reward. To attain these rewards, participants made choices during the first stage, which probabilistically determined the fractals presented during the second stage. B Illustration of trial sequences, showing a choice made in the first stage, followed by feedback, and a second-stage selection that was followed by a reward (1 play pound). C A fractal to response key pairing was allocated randomly in each trial. Panel (C) illustrates a trial sequence, in which the same fractals were selected as in panel (B), but now with different effectors. Although fractal identity predicted relevant outcomes (second-stage fractals, and reward), the position of the fractal and the response key used to report a selection were always outcome-irrelevant. Outcome-irrelevant learning was inferred from a participant’s tendency to assign value to response keys despite their irrelevance to any individual decision (see Fig. 3 for outcome-irrelevant estimate plots). Model-based control was estimated as the ability to select a first-stage action based on the task’s transition probability and subjective action values of second-stage fractals [14, 31] (see Supplementary Fig. S3 for model-based estimate plots).