English Español Français Deutsch
← Back to the list
Hallucination and factuality Methodologists Directly upstream of ai100's methodology

Yarin Gal

Quantifying when language models don't know what they're saying.

Gal worked out the foundations of Bayesian uncertainty in deep neural networks while that was still a small literature, and has spent the past few years pointing the same instrument at LLMs. His 2024 Nature paper on semantic-entropy hallucination detection is the throughline made concrete: a fluent answer and a true one are different things, and the gap is something you can measure from the outside rather than ask the model to self-report. The method does not require labeled ground truth or access to model internals, which is the only kind of method that actually applies to a closed commercial LLM.

Worth following when
you need to take a language-model output and ask "should this be trusted?" — and you want an answer that doesn't depend on the model agreeing.
Topics
semantic uncertainty in LLMs; hallucination detection without labeled ground truth; how AI safety institutes translate uncertainty research into oversight.
Key works
"Dropout as a Bayesian approximation" (2016); "Detecting hallucinations in LLMs using semantic entropy" (Nature, 2024); ongoing Oxford OATML group output.