Supplemental Material for Detecting Social Biases Using Mental State Inference

Supplementary materials for "Detecting social biases using mental state inference" contain additional details about the models, and methods and analyses for the experiments. Computational ModelWe use the same four models, one main ("Full") model and three alternative models, to generate predictions for all experiments.Here we provide specifications and implementation details for each model.Full model code is available at https://osf.io/kprz3/?view only= 25d06dd71efb4b8699ab151ab40e4bc5.All models were implemented in WebPPL, a probabilistic programming language for generative models. Main modelA single forward pass of our main generative model comprises the following steps:1. Sample the actor's true competence c true from a symmetric Beta distribution Beta(5, 5).2. Compute the actor's probability of success on each task (easy, medium, hard), based on the sample competence value c true and the difficulty value of each task, which are fixed to d easy = .1,d medium = .3,and d hard = .7.On a single attempt, the actor will get lucky and succeed with probability .1,get unlucky and fail with probability .1, or succeed with probability 1/ (1 + exp( (c trued)).This is a standard logit transformation with temperature parameter , which we set equal to 4 based on maximum likelihood estimation from an earlier pilot study.