A Unified Definition of Hallucination: It's The World Model, Stupid!

The paper presents a unified definition of hallucination in language models, characterizing it as inaccurate internal world modeling observable to users, such as contradicting established facts. This framework allows for a clearer evaluation of hallucinations by identifying the reference world model and distinguishing them from other types of errors. Additionally, it introduces HalluWorld, a benchmark designed to rigorously test model hallucinations against specified reference world models, which is crucial for practitioners aiming to enhance the reliability of LLMs.