Preprint: Increased Reasoning Compute Causes LLMs to Cancel Pragmatic Implicatures of 'Some' and 'Or'

A newly released arXiv preprint indicates that increasing the thinking budget in large reasoning models leads to overly literal interpretations of everyday logical words. When a speaker says "some students passed," human listeners typically infer that not all did. However, the study found that as internal reasoning steps increase, most tested open-source reasoning models and GPT-5 cancel this "conversational implicature," adopting a strict logical definition where "some" means only "at least one," without implying "not all."

The research team conducted experiments by keeping model weights fixed and varying only the amount of thinking time, recording which interpretation the model chose. They observed that models initially entertain the enriched pragmatic reading but cancel it in later reasoning steps. In contrast, context-driven inferences like indirect answers improved with more thinking. The authors propose an exploratory remedy: answering from the early stage of reasoning before the implicature is cancelled can recover these meanings without sacrificing literal accuracy.