Signal Toward a Gricean Retreat: Probing LLMs for Knowledge Boundaries and Referent Specificity

Summary

When asked about entities beyond their knowledge boundary, large language models routinely fabricate plausible-sounding details rather than retreating to safer, more general claims, a core driver of hallucination. The authors frame this failure through a Gricean lens. A cooperative human speaker who is unsure about a specific referent retreats up a hierarchy of specificity, trading informativeness for truthfulness. For instance, saying a European city rather than guessing the wrong specific one. The paper probes whether current LLMs have the internal ingredients, knowledge-boundary awareness and referent-specificity control, needed for this kind of graceful epistemic retreat instead of fabrication, which bears directly on hallucination mitigation and trustworthy AI deployment.