MULTI-TIMESCALE DYNAMICS IN PREFRONTAL CORTEX UNDERLYING ALGORITHMIC FLEXIBILITY DURING PROBABILISTIC DECISION-MAKING

Perceptual decisions depend not only on immediate sensory evidence but also on statistical regularities accumulated over time. How neural populations integrate slowly evolving environmental statistics with rapid sensory and motor processes remains poorly understood. Our recent behavioural work shows that rats exhibit pronounced algorithmic heterogeneity in probabilistic categorisation tasks: despite match performance, animals alternate between learning full stimulus-category distributions (generative strategies) and simpler discriminative boundary-estimation strategies that ignore within-category structure[Menichini*,Pajot-Moric*,Low* et al 2025].
To identify the neural mechanism underlying this cognitive flexibility, we recorded large-scale neural activity from medial prefrontal cortex(mPFC) while rats performed a two-alternative forced-choice task with uncued changes in sensory statistics. We analyse these data using unsupervised, non-stationary dynamical systems approaches to identify neural states and trajectories without imposing task labels or behavioural models. This framework allows us to dissociate neural dynamics operating at multiple timescales: (i) fast, transient dynamics related to sensory processing and action selection intra-trial; (ii) persistent dynamics reflecting block-level changes in stimulus prior distributions; (iii) slower, session-level dynamics associated with shifts between generative and discriminative learning algorithms.
We hypothesise that mPFC population activity occupy distinct subspaces encoding task structure across these timescales, with persistent dynamics functioning as neural priors, tracking block-level probabilities before stimulus onset. Moreover, generative learning epochs should be associated with higher dimensional, persistent dynamics encoding distributional structure and uncertainty, whereas boundary-estimation epochs should exhibit lower-dimensional, action-aligned dynamics.
These results will reveal how mPFC in rats reconfigures its dynamics to control which learning algorithm governs behaviour, independently of external task variables.