A new arXiv preprint, 'Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally,' investigates a subtle question: what does extra thinking do to everyday pragmatic inferences? In ordinary speech, saying 'some students passed' often implies that not all did. The paper asks whether reasoning models, which now let users set how long they think before answering, preserve or lose such implicatures.

The abstract frames the problem but does not give full results in the available text. The title, however, suggests a clear direction: with more thinking, models may 'entertain' the stronger interpretation and then cancel it, ending up with a literal reading of 'some' as 'at least some' and 'or' as inclusive. If that holds, longer reasoning could make models less pragmatically natural, not more.

Because only the abstract is available, the article cannot confirm effect sizes or experimental details. The work is nonetheless notable as one of the first to connect adjustable inference-time compute with scalar implicature, a core feature of human communication.