Humans can seamlessly infer what other people like, based onwhat they do. Broadly, two types of accounts have beenproposed to explain different aspects of this ability. A firstaccount focuses on inferences from spatial information:agents choose and move towards things they like. A secondaccount focuses on inferences from statistical information:uncommon choices reveal preferences more clearly comparedto common choices. Here we argue that these two kinds ofinferences can be explained by the assumption that agentsmaximize utilities. We test this idea in a task where adultparticipants infer an agent’s preferences using a combinationof spatial and statistical information. We show that our modelpredicts human answers with higher accuracy than a set ofplausible alternative models.