Figure 2.

Computational model results. a, Model simulations of the likelihood of sharing with a partner on a given trial for one randomly selected participant's experimental data. We compared the ability of the social value model (bottom) to explain collaborative behavior with a standard expected value model (top), which assumes that participants maximize expected value based solely on self-interested financial value and a 50% reinforcement rate, and an expectation-learning model incorporating strong priors, in which participants update their beliefs about partner reciprocation from both positive and negative outcomes (middle). b, Average model fits penalizing for the number of free parameters using the Akaike Information Criteria (AIC). The social value model provided the best fit to participant behavior (n = 26). c, Average social value bonus (theta * normalized trustworthiness ratings) for each partner. ***p < 0.0001; **p < 0.001; *p < 0.05 (±SEM).