Understanding how information is distributed over time is a core problem in natural language modeling, according to a new paper posted on arXiv. The authors argue that existing empirical work has largely relied on post-hoc analysis of information-theoretic metrics such as surprisal, and that this leaves a gap: little is known about the arrival time of information-bearing events themselves.
The paper's title points to a modeling approach that treats the timing of information as the object of study, rather than treating surprisal as a static property of words or sentences. The abstract does not provide details of the proposed model or experimental results, so the contribution is best understood as a framing and motivation for a new line of inquiry.
If the approach holds up, it could shift how researchers evaluate language models and cognitive models of reading, moving from aggregate predictability scores to a more dynamic account of when information arrives. But without further details from the full paper, those implications remain speculative.