MEMORY-DEPENDENT BIASES REVEAL HIDDEN PRIORS IN A TWO-STEP AUDITORY CATEGORISATION TASK

In standard one-step perceptual categorisation tasks, animals directly map a sensory stimulus onto a categorical action, such that perceptual inference and motor planning occur simultaneously. This tight coupling makes it difficult to isolate intermediate sensory representations and latent decision variables. We developed a novel two-step working memory (WM) based sound categorisation paradigm that temporally dissociates perceptual inference from action selection. In this task, head-fixed mice categorise a continuous auditory stimulus onto an abstract category, followed by a delayed cue that determines how the category maps onto a motor response. By manipulating the delay interval, we probe how internal representations evolve under increasing WM demands. Mice reliably learned this cognitively demanding task and performed stably in sessions with randomly interleaved trial types. Using logistic regression models incorporating sensory evidence, task rules, and trial history, we revealed consistent, subject-specific decision biases that could not be explained by disengagement or lapse behaviour and instead reflect learned priors shaping categorical inference. Increasing the WM delay did not uniformly impair performance; rather, it selectively amplified these biases and shifted reliance away from the rule-defining stimulus-cue interaction towards single task cues. Together, these results demonstrate that task structure and memory demands critically shape how learned priors are expressed in behaviour. Apparent optimality in one-step tasks can mask systemic biases that emerge only under memory and rule constraints and can be studied when inference and action are decoupled. This two-step paradigm provides a powerful framework for investigating how internal representations are maintained and transformed during context-dependent decision-making.