Next-Token Prediction Enables Scalable Learning of Sleep Physiology

Foundation models offer a promising route to compress multi-modal physiological signals into better measures of human health, with broad applications across sleep medicine, cardiology and neurology. Existing models are typically trained with contrastive or masked-reconstruction objectives, both of which have known shortcomings on stochastic, continuous signals. In this work, we show that next-token prediction is a simple and scalable alternative. We develop Hypnos, a multi-modal sleep foundation model trained via next-token prediction over residual-vector-quantization tokens drawn from eight sensing modalities of overnight polysomnography recordings (e.g.\ EEG, ECG, respiratory signals). A large auto-regressive RQ-Transformer jointly predicts the next token across all modalities in parallel, with a novel modality-masking strategy enabling generalisation to subsets of modalities during inference. Using over 20{,}000 overnight recordings drawn from nine public datasets, we find that both next-token perplexity and downstream probing performance continue to improve with model scale. Hypnos matches or exceeds prior sleep foundation models and strong supervised baselines on sleep stage classification across in-domain and held-out test sets.