Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally

Entertain, then Cancel. How Thinking Makes Reasoning Models Read Some and Or Literally. By Zhizhuo Kou and six others, from the Hong Kong University of Science and Technology, the Hong Kong University of Science and Technology, Guangzhou, and one other institution. Published on the fifteenth of October. Its code repository is worked out: about one hundred source files, with a readme and a licence.

When someone says that some students passed, listeners infer that not all of them did. Reasoning models now let users choose how long a model thinks before answering, and nobody has known what extra thinking does to everyday inferences of this kind. The finding here is that thinking makes models read logical words as logic. With more thought, they cancel the implicature of some, meaning not all, and of or, meaning not both, and fall back on the strict logical reading.

The study isolates thinking itself. A model's weights stay fixed; only the amount it thinks is changed, and the reading it chooses is recorded. A second tool cuts the model's own reasoning off at successive points and asks for an answer, which shows where along the trace the reading changes.

The cancellation appears in most of the tested open reasoning models, and in GPT-5. The trace explains it: models first entertain the enriched reading, and cancel it later. That suggests a remedy, which the authors describe as exploratory and applied at inference time. Answering from the early reading recovers the implicatures without losing literal accuracy.

The effect is not that thinking harms every inference. Context-driven inferences, such as understanding indirect answers, improve with more thinking. The loss is specific to implicatures carried by logical words like some and or. The pattern holds in most, not all, of the open models tested, and the remedy is offered as exploratory rather than settled.

The authors draw a lesson for evaluation: reasoning models should be measured on how they interpret, not only on how accurate they are. A model can be logically correct and still miss what a speaker meant, and with these words, thinking longer is what takes it there.