Pragmatic Reasoning in Language Models

This field studies how language models interpret what people mean beyond literal wording, including implicatures, indirect requests, ambiguity, and culturally shaped conventions. Researchers use theory-based tests such as Rational Speech Act models, controlled benchmarks, and interactive dialogue tasks to examine how models track speaker beliefs, common ground, and conversational context. Results show that models can often recognize or judge pragmatic meanings more reliably than they can use them when generating responses, and explicit prompts or targeted training sometimes help. A central challenge is making these abilities robust across speakers, languages, and real interactions, where fluent or accurate answers do not necessarily show that a model understood the user’s intent.