Extended Data Fig. 1. Additional model simulations and illustration of the model space.

a, Simulation of model Q1 under partial feedback. Same conventions as in Fig. 2. The simple Q-learning models (Q1/Q2) can only learn about the extreme items (here, A and H) under partial feedback. b, Choice matrices predicted by pair-level learning without (left, model P) or with associative recall of ‘linking’ pair relationships (right, model Pi). Choice behaviour was simulated with a pair-level learning rate γ=1. Associative recall in model Pi (right) was enabled by additionally lowering parameter λ to 1 (see Methods for details). c, Schematic overview of the model space. See Methods, Models.