In the context of Large Language Models, a "hallucination" is defined as a confident response that is factually incorrect or unsupported by the prompt. However, from the perspective of the neural network, it is simply navigating the latent space based on its training distribution.

The Mechanism of Creativity

Human creativity often involves combining disparate ideas to form something new. When an LLM does this and produces a novel solution to a coding problem, we call it "zero-shot reasoning." When it does the exact same thing but produces a fake legal citation, we call it a "hallucination."

The algorithmic process behind invention and hallucination is identical. The only difference is whether the output aligns with our external, objective reality.

OutcomeProcessValue
Fake CitationLatent Space InterpolationNegative (Deceptive)
Novel MetaphorLatent Space InterpolationPositive (Creative)

FAQ: Hallucinations

Can we fix hallucinations entirely?
Likely not without fundamentally altering how LLMs work. To eliminate hallucination is to eliminate the model's ability to synthesize novel ideas.