The formal framework is slated for presentation at ICML.
We keep saying LLMs "hallucinate." But what does that actually mean?
In our new position paper, we argue hallucination isn't just "wrong facts." It's inaccurate internal world modeling.
We formalize this precisely in a unified definition to appear at #ICML2026 (@icmlconf)👇
No Digg Deeper questions have been answered for this story yet.
A while ago, @roydanroy asked: "Do we have a definition of hallucination yet?" This paper is our attempt at one.
We argue that hallucination isn't just a surface-level language problem, but a failure of the system's internal world representation to align with a reference world.
If LLMs are world-modeling systems (as many argue), then hallucination is a world-modeling error.
Inspired by the idea that intelligent systems build internal world models (@ylecun).
