Abstract
<p>Bayesian accounts of time perception propose that temporal estimates arise from integrating noisy sensory evidence with prior knowledge about interval statistics. However, the extent to which observers adapt their assumptions about temporal continuity to the structure of the environment remains unclear. We tested whether people flexibly adjust these assumptions during interval reproduction. Across two experiments, participants reproduced intervals drawn either from a randomised sequence or from an autocorrelated random-walk sequence. We also manipulated object context to bias perception toward a single persisting object versus multiple discrete objects, reasoning that temporal continuity may be more likely to be inferred when successive intervals are associated with a common source. Behaviour was evaluated using Bayesian model comparison among a stationary distribution-learning model, a non-stationary tracking model, and a two-state hybrid model capturing intermediate assumptions about temporal continuity. Participants exhibited robust central tendency biases across conditions. However, model fits and Akaike weights provided little evidence that observers adopted stronger assumptions of temporal continuity when intervals were autocorrelated or when object context favoured single-object tracking. Instead, responses were best explained by learning under an effectively stationary prior, with two-state fits indicating only limited contributions of temporal continuity. These findings indicate limited flexibility in prior updating for temporal perception: rather than adapting their assumptions to local temporal structure, observers appear to rely primarily on stable estimates of interval statistics within the session.</p>