Abstract
Vision can be viewed as a continuous information processing, yet its underlying system properties have not been fully understood. Studies of visual serial dependence suggest that current perception is often biased by the preceding stimuli, raising the possibility of Markov-like processing-where only the previous state (not the ones before) affects the current one. In the current study, participants rated faces on two of three traits (attractiveness, trustworthiness, and dominance), presented in randomized sequences so each rating could be preceded by the same or a different trait. This design allowed us to examine how prior input (the face) and prior output (the perception) influence current judgment. Using derivative of Gaussian, Markov chain, and linear mixed-effects modeling, we found that serial dependence was disrupted-and both memoryless property and Markov assumptions were violated-when alternating between two traits for attractiveness and dominance but not under other conditions. These findings suggest that different facets of (presumably) the same visual computation can exhibit asymmetrical system properties. More broadly, our work shows how serial dependence can serve as a powerful tool to probe the underlying rules by which the visual system integrates past and present information.