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

Amir Globerson

Using one language model as an adversarial interrogator of another to surface factual errors that neither could find alone.

The 2023 "LM vs LM" paper from Globerson's group set up a cross-examination protocol: one model produces a claim, a second model asks follow-up questions designed to probe for inconsistencies, and the original claim is flagged if the follow-ups force a contradiction. The setup gives the field something it had been missing — a factuality-detection method that doesn't require labeled ground truth, doesn't require model internals, and doesn't require trusting either model on its own. Globerson's longer arc is in machine-learning theory, which shapes how the work reads: LLM evaluation as a problem about adversarial sampling and information geometry, rather than as a problem about prompt engineering.

Worth following when
you want factuality detection methodology grounded in theoretical first principles rather than empirical recipe-hunting.
Topics
adversarial LM-vs-LM evaluation protocols; factuality detection without labeled data; learning theory as a lens on LLM behavior.
Key works
"LM vs LM: Detecting Factual Errors via Cross-Examination" (2023); broader ML-theory work informing the evaluation methodology.