A Unified Definition of Hallucination: It's The World Model, Stupid!
A new paper proposes a unified definition of AI hallucination, framing it as inaccurate internal world modeling observable to the user. This definition aims to clarify evaluation methodologies by requiring explicit reference to a 'world model' and distinguishing true hallucinations from other error types. The authors connect this framework to HalluWorld, a benchmark designed to test model hallucinations by specifying reference world models. AI
IMPACT Provides a clearer framework for understanding and evaluating AI hallucinations, potentially leading to more robust mitigation strategies.