English Español Français Deutsch
← Back to the list
Foundational figures Methodologists Directly upstream of ai100's methodology

Roi Reichart

What "domain" means for a language model — when its training distribution stops matching its deployment context — and how to evaluate that mismatch rigorously.

Reichart's longer research line has been domain adaptation in NLP: when a model trained on news articles is deployed on legal text, what specifically breaks, and how do you measure the breakage in a way that doesn't depend on the model itself being legible. As Co-Editor-in-Chief of the Transactions of the Association for Computational Linguistics, he has institutional influence on what the field accepts as rigorous evaluation methodology — TACL is one of the venues where evaluation papers either survive review or don't, and the standards there shape what subsequent work has to clear. For ai100, which audits AI engines on queries the engines may not have been optimized for, the domain-adaptation literature is precisely the methodological backing.

Worth following when
you want to evaluate a language model's behavior on inputs that may differ from its training distribution, and want the methodology that takes that mismatch seriously.
Topics
domain adaptation methodology for language models; active learning under distribution shift; the editorial standards of TACL as a venue.
Key works
body of work on NLP domain adaptation (2010s onward, Technion); long-term Co-Editor-in-Chief role at Transactions of the ACL; active-learning publications applied to language tasks.