Extra thinking makes reasoning models cancel some and or implicatures
Original title: Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally
Models first entertain the enriched reading, then cancel it; early answers recover implicatures without losing literal accuracy.
Researchers kept model weights fixed and varied only how long models think before answering everyday inferences. Extra thinking made most tested open reasoning models and GPT-5 cancel implicatures of "some" (not all) and "or" (not both), while context-driven inferences such as indirect answers improved. Trace cuts show models entertain the enriched reading first and cancel it later. Answering from that early reading recovered implicatures in an exploratory test.
Unconfirmed: The early-answer remedy is described as exploratory, not a validated fix.
- Only thinking length changed; weights stayed fixed.
- GPT-5 and most tested open reasoning models showed the cancellation.
- Context-driven inferences improved as thinking increased.
- Authors say evaluations should measure interpretation, not only accuracy.
From the original article
1 min read at huggingface.co
When a speaker says ``some students passed'', listeners infer that not all of them did. Reasoning models now let users decide how long a model thinks before it answers, but we do not know what extra thinking does to such everyday inferences. We study this directly: we keep a model's weights fixed, change only how much it thinks, and record which reading it chooses.
By Zhizhuo Kou, Jian Yang, Pengfei Hao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.