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

One-line takeaway

Thinking longer makes reasoning models (GPT-5 included) read 'some' and 'or' literally, cancelling everyday implicatures; traces show models first entertain the enriched reading then cancel it, and answering from the early reading recovers it.

研究发现推理模型想得越多，越把 some、or 读成纯逻辑义，丢掉“不是全部”“不是两者都”的日常含义，GPT-5 也一样；轨迹显示模型先想到含义再推翻，用早期读法作答可挽回。