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

Flexible decision-making in uncertain environments requires inferring latent structure and selecting behaviourally relevant information. Here, we tested the hypothesis that internal models support this process by compressing high-dimensional input into lower-dimensional, goal-relevant subspaces. Human participants performed a serial reversal learning task requiring dynamic inference of the currently relevant stimulus dimension(s). Behavioural data showed rapid adaptation to context switches, with improving accuracy and faster responses over trials. A hidden state inference (HSI) model, which maintains probabilistic beliefs over latent contexts, best explained behaviour, outperforming multiple reinforcement learning variants in predicting both choices and inferred context. Eye-tracking revealed that gaze selectively converged on task-relevant features over time, with reduced attentional entropy and increased Uixation within relevant regions, hallmarks of belief-guided attention. Crucially, gaze patterns were tightly linked to model dynamics: attentional focus increased when belief updating succeeded and decreased after trials where the model signalled high prediction errors. Guided by these results, we used magnetoencephalography (MEG) to understand the goal-relevant transformations within subspaces and found that neural representations expanded along relevant dimensions and compressed along irrelevant ones, with effects emerging just before decisions. These modulations of space correlated with reaction time, such that participants responded faster when the distance along relevant axes increased. Strikingly, these transformations preceded feedback-locked frontal theta responses modulated by prediction error, suggesting that internal models proactively reconUigure cortical state space in anticipation of learning. Together, these Uindings highlight a mechanism by which internal models, attention, and neural representations interact to support abstraction and adaptive decision-making.