Hidden state inference or continuous belief updating during a dynamic visuomotor skill

Adaptive sensorimotor behaviour requires individuals to flexibly update their beliefs in response to changes in environmental context. Theories of predictive processing propose that such flexibility arises from hierarchical inference, where higher-level beliefs about hidden states shape lower-level perceptual and motor predictions. Here, we test whether behaviour in a naturalistic interception task is better explained by continuous belief updating or by more abrupt shifts driven by state inference, consistent with hierarchical Bayesian learning. Twenty-three participants completed 160 trials of a VR-based interceptive task across two sessions. Participants attempted to return balls with varying bounciness, probabilistically cued by ball colour and a contextual cue (court wall colour) that reversed periodically. Gaze pitch angle prior to bounce was used as an index of perceptual inference and modelled using Bayesian and reinforcement learning frameworks. Bayesian models (the Hierarchical Gaussian Filter) were found to outperform associative learning models in predicting participant behaviour. However, although a contextual state-inference model better anticipated the true structure of the task, participants’ gaze behaviour was better explained by continuous belief updating rather than discrete state switching. Overall, our findings suggest that while participants updated beliefs in a Bayesian manner, their behaviour fell short of normative optimality, possibly due to limits on causal inference and model construction. Rather than inferring the latent structure of the environment, participants relied on simpler, continuous updating, highlighting how probabilistic reasoning operates under bounded resources. This finding may reflect the specific demands of dynamic sensorimotor control, where the brain prioritises flexible, real-time updating over explicit structural inference to support fluid, adaptive action.