Across four experiments, we look at whether adults andchildren can represent the amount of information needed todistinguish different populations in the context of an intuitivestatistical reasoning task requiring metacognitive monitoringand control. Consistent with a ground truth model ofinformation gain, adults (N=60) modulated their informationgathering with respect to the difficulty of the discriminationproblem. Adults also adjusted their confidence thresholddepending on task difficulty, allowing for more uncertainjudgments when the discrimination was more difficult orgathering data was more costly (Experiments 1 and 2). In asimplified version of the task, children (N = 42, M = 7.3years, range: 5.0-9.0) were also able to distinguish easy anddifficult discrimination problems and judge that they neededmore information to solve harder problems (Experiments 3and 4).

From minimal observable action, humans automatically make fast, intuitive judgments about what other people think,want, and feel (Heider & Simmel, 1944). Even when no agent is visible, children can infer the presence of intentionalagents based on the environmental traces that only agents could leave behind (Saxe et al., 2005; Newman et al., 2010).Here we show that four- to six-year-olds can go beyond inferring the presence of an agent to matching an agents mentalstate with the trace they left behind. Participants (N = 35, M: 5.6 years, range:4.0 6.8 years) saw pairs of dresser drawerswith different numbers and orientations of open drawers, and were asked to match one of the static scenes to an agentsknowledge state (whether the agent wasnt searching at all but was just playing, knew exactly where an object was hidden,knew the approximate location, had no idea where it was hidden, or at first didnt know and then remembered). We comparechildrens performance to a formal model, in which we build upon classical models of Bayesian Theory of Mind that treatmental state inferences as a form of inverse planning; here we extend those models to consider cases where the behavioris not observed but must be inferred from the structure of the environment.

Previous work has argued that young children do not answercounterfactual questions (e.g. “what would have happened?”)by constructing simulations of alternative possibilities in theway adults do. Here, we propose that children can engage insimulation when answering these questions, but considerdifferent counterfactual possibilities than adults. While mostprevious research has relied on narrative stimuli, we use causalperception events, which are understood even in infancy. InExperiment 1, we replicate earlier findings that childrenstruggle with counterfactual reasoning, but show that they arecapable of conducting the required simulations in a predictiontask. In Experiment 2, we use a novel multiple-choice methodthat allows us to study not only when children get it right, butalso how they get it wrong. We find evidence that 4-year-oldsengage in simulation, but preserve only some features of whatactually happened and not others.