Hidden-state inference aligns attention and neural representational geometry during flexible behaviour

Abstract

Flexible decision-making in uncertain environments requires inferring latent task states and prioritising behaviourally relevant information. We tested the hypothesis that this process is associated with systematic changes in attention and neural representational structure. 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, and a hidden-state inference model best explained these adaptations, outperforming multiple reinforcement-learning variants in predicting both choices and inferred contexts. Oculomotor behaviour provided converging evidence for selective attention, with gaze increasingly concentrated on relevant features, tracking trial-by-trial belief updates, and transiently broadening following high prediction errors. Guided by these results, we used magnetoencephalography to examine how latent-state inference was reflected in neural activity. Neural representations encoded the currently inferred context and were modulated by recent prediction errors, with larger prediction errors associated with weaker context representations on subsequent trials. Representational analyses further revealed that successful performance was associated with a systematic reorganisation of neural geometry, characterised by selective amplification of task-relevant representational distances and increased separation of behaviourally relevant stimulus values. Together, our findings demonstrate that latent-state inference, attention, and neural representational geometry are tightly coordinated during flexible decision-making, providing a systems-level account of how behaviourally relevant information is selectively prioritised and represented to support adaptive behaviour.