Individualised dynamic internal representations from response times

Internal models are central to understand how human behaviour is adapted to the statistics of, potentially limited, environmental data. Such internal models contribute to rich and flexible inferences and thus adapt to varying task demands. However, the right internal model is not available for observers, instead approximate and transient internal models are recruited. To understand learning and momentary inferences, we need tools to characterise these approximate, yet rich and potentially dynamic models through behaviour. We used a combination of non-parametric Bayesian methods and probabilistic programming to infer individualised internal models from human response times in an implicit visuomotor learning task. Using this Cognitive Tomography approach we predict response times on a trial-by-trial basis and validate the internal model by showing its invariance across tasks and sensitivity to stimulus statistics. By tracking the performance of participants for multiple days, individual learning curves revealed transient subjective internal models and pronounced inductive biases.