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

The paper studies how extended thinking affects scalar implicatures, such as reading 'some' as 'not all' and 'or' as 'not both'. Holding model weights fixed and varying only thinking length, the authors find that thinking tends to cancel these enriched readings in most tested open reasoning models and in GPT-5, while context-driven inferences like indirect answers improve. Reasoning traces suggest models first entertain the enriched reading and cancel it later. Answering from this early reading is described as an exploratory remedy that recovers the implicatures without losing literal accuracy.

Why it matters

Reasoning models may become more literal in everyday interpretation as they think longer, so evaluations should test interpretation alongside accuracy.

Read with care

Source claims not independently checked; the inference-time remedy is described by the paper as exploratory and its generality is unverified.