LLMs Fabricate Details on Unknown Entities; Study Proposes Gricean Retreat

A recent paper on arXiv (2608.13484) delves into the reasons behind large language models (LLMs) generating believable yet inaccurate information when queried about unfamiliar entities, instead of opting for safer, broader statements. The authors apply a Gricean perspective, arguing that a cooperative speaker unsure of a reference should choose a less specific answer, prioritizing accuracy over detail. By utilizing a T-REx-based benchmark that assesses both entity familiarity and referent specificity, the researchers examine whether models recognize their knowledge limits and predict the specificity of the information they generate. Their findings reveal that while both signals exist, models still favor specific references even when the entity is unknown, ignoring correct generic options. This study enhances the understanding of LLM hallucinations and proposes ways to improve model reliability and calibration. The paper was introduced as a cross-type submission on arXiv, with its abstract accessible via the provided URL.