LLMs Struggle with Communicative Belief Updates and Implicature Cancellation
July 29, 2026
Evaluation via a new expert-annotated dataset reveals that LLMs lag behind humans in recognizing unspoken beliefs and understanding implicature cancellation. Findings suggest LLM success in belief updates often relies on prior knowledge rather than true pragmatic reasoning.
HOW THIS AFFECTS YOU
●
researcherThis highlights a specific reasoning gap in current LLM architectures regarding pragmatic linguistics.
●
policyUnderstanding these communicative failures is critical for assessing the reliability of LLM-human interactions.