We present incremental alternative sampling (IAS), a computational account of incremental language comprehension in which the comprehender continually generates, evaluates, and updates predictions about alternative continuations of partial linguistic input. IAS is related to information-theoretic approaches to predictive processing, including surprisal, but makes explicit the role of variable prediction horizons and multiple levels of linguistic representation. Within this framework, we define incremental information value as the representational distance between predictions generated before and after observing a unit, thereby quantifying the change in predictive uncertainty induced by that unit and, by extension, its predictability and associated processing effort. We operationalise this measure using Transformer-based language models — both to sample alternative continuations spanning variable prediction horizons and to derive representations along a continuum of linguistic levels. Language model estimates of incremental information value significantly predict a wide range of word-level neural and behavioural responses to incremental stimuli, including cloze probabilities, predictability ratings, eye-tracking measures, self-paced reading times, and ERP amplitudes. IAS also offers insight into the predictive strategies underlying these responses. Distinct psycholinguistic measures are best captured by different combinations of temporal and representational resolution, and a model that integrates multiple resolutions outperforms surprisal for most measures, with particularly strong effects for cloze probability, N400 amplitudes, and for self-paced reading times in naturalistic, multi-sentence stimuli. Beyond modelling human comprehension, IAS provides a principled lens on predictive mechanisms in artificial language processors, showing that next-word prediction in Transformer language models implicitly encodes uncertainty over longer temporal horizons and across levels of representation.