When a speaker says “some students passed”, listeners infer that not all of them did. Recent reasoning models let users set how many thinking steps the model takes before answering, yet it is unclear how extra thinking influences such everyday inferences. We study this directly: we keep the model’s weights fixed, vary only the amount of thinking, and record which interpretation the model adopts. A second tool cuts off the model’s own reasoning at successive points and asks for an answer, revealing where the interpretation switches. Experiments show that longer thinking makes models treat logical words as strict logical operators: in most open‑source reasoning models and in GPT‑5, the implicature of some (not all) and or (not both) is cancelled, while context‑driven inferences such as indirect answers improve. Tracing the process shows that models first “entertain” a richer reading and later cancel it. Answering from that early stage—an exploratory inference‑time remedy—recovers the implicatures without losing literal accuracy. Therefore, evaluations of reasoning models should measure how they interpret, not only how accurate their answers are.
Review