Dynamic Sensorimotor Behavior Reveals Bayesian Belief Updating

In everyday behavior, humans react and respond to their environment. Over time, humans learn to anticipate what events typically come next, guiding their perceptions and adaptive behaviors. It has been suggested that to accurately anticipate events, humans form internal models that represent how the world works and how actions relate to specific outcomes (Wolpert et al., 1995). These models reflect perceptions of the world and give rise to predictions about events and opportunities for action that exist within given environments. As individuals gain experience, internal models are continuously updated. When opportunities for action arise, responses are considered and are weighted by contextual information. Feedback from the chosen behavior then updates the internal model, and the cyclical nature of sensorimotor processing continues. Overall, these internal models are adaptive, guiding perception and action in dynamic environments. Important to internal models of the world are the probabilities of behavioral outcomes. For example, licensed drivers show a learned sensitivity for identifying pedestrians who seem to intend to cross into the road, but only when driving (Britt et al., 2025). The driving context signals to the driver that certain events (e.g., pedestrians entering the roadway) are more relevant to their ongoing behavior and warrant attentional priority. According to Bayesian frameworks, humans learn in different contexts by continuously updating their beliefs or prior expectations with new evidence. Past and present information is weighted according to their respective reliabilities (Körding and Wolpert, 2004). If certain information is deemed more reliable, it will exert greater influence on future behavior. Prominent models of Bayesian learning are hierarchical, where predictions are made and organized across levels. Low-level sensory predictions are formed from immediate sensory information, while high-level sensory predictions are governed by more abstract thoughts and properties, such as context cues (Friston et al., … Correspondence should be addressed to Noah Britt at brittn{at}mcmaster.ca.