Inference over hidden contexts shapes the geometry of conceptual knowledge for flexible behaviour

Flexible decision-making in uncertain environments requires inferring the relevant context and focusing on behaviourally relevant information. We tested the hypothesis that the brain supports this process by compressing high-dimensional input into lower-dimensional, goal-relevant subspaces. Human participants performed a serial reversal learning task that required identifying which stimulus dimension(s) were currently relevant. Behaviourally, participants rapidly adapted to context switches. A hidden-state inference model best explained these adaptations, outperforming multiple reinforcement-learning variants in predicting both choices and inferred contexts. Oculomotor behaviour corroborated this account where relevant features were selectively attended to, with attentional focus tracking the model’s belief updates and decreasing after high prediction errors. Guided by these results, we used magnetoencephalography to examine how cortical population activity is reorganised during inference. Dimensionality and geometry analyses showed that neural state space expanded with the inferred complexity of the task and that 2D contexts were represented as two near-orthogonal axes within a shared planar manifold. Neural activity expanded along relevant as compared to irrelevant dimensions, with these transformations predicting faster reaction times and emerging just before decisions. Finally, these representational changes preceded feedback-evoked frontal theta responses sensitive to prediction error, suggesting that internal models proactively reshape cortical state space to facilitate upcoming learning. Together, our findings reveal how latent-state inference, attention, and dynamic representational geometry interact to support abstraction and adaptive decision-making.