Metacognition & Hallucination

The Problem

Large language models routinely produce confident, fluent, and false statements. This "hallucination" problem is tied to metacognition: models often express equal certainty whether they're right or wrong, lack a reliable internal "I don't know" signal, and distinguishing a grounded answer from a fabricated one remains an open problem.

What We're Working On

• Metacognition in LLMs — Whether and how models represent the limits of their own knowledge.
• Hallucination detection — Flagging fabricated or ungrounded outputs, including from a model's internal states.
• Honesty and calibration — Improving how reliably expressed confidence tracks the truth.