Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally
Extra thinking led most tested reasoning models to drop the everyday inference that “some” means “not all.”
Researchers held model weights fixed and varied thinking time, finding this effect in most tested open reasoning models and GPT-5. Models also dropped the inference that “or” means “not both,” while improving on context-driven inferences such as indirect answers. Cutting reasoning traces at successive points showed models considering the implied meaning early, then canceling it later. An exploratory method that answered from the earlier reading recovered those implications without reducing literal accuracy.
HF Daily Papers' note
Researchers held model weights fixed and varied thinking time, finding this effect in most tested open reasoning models and GPT-5. Models also dropped the inference that “or” means “not both,” while improving on context-driven inferences such as indirect answers. Cutting reasoning traces at successive points showed models considering the implied meaning early, then canceling it later. An exploratory method that answered from the earlier reading recovered those implications without reducing literal accuracy.
HF Daily Papers' note
score 5