LLMs Struggle with Pragmatic Reasoning under Epistemic Asymmetry
July 14, 2026
LLMs fail to consistently follow Grice's cooperative principle when working in multi-party pipelines with partial information. The study evaluates prompting and post-training strategies, finding that models struggle to act as effective listeners or speakers when information is asymmetric between collaborators.
HOW THIS AFFECTS YOU
●
builderBe cautious when deploying LLMs in multi-agent workflows involving asymmetric data access.
●
researcherYou should investigate more robust post-training methods for pragmatic reasoning.