Rational Analysis of Language Models 

Join us for a talk by Andrew Lampinen, Staff Research Scientist at DeepMind. This talk is part of the Kempner Seminar Series, a research-level seminar series on recent advances in the field.

There has been substantial debate about the capabilities of language models—which aspects of language they can acquire, whether they can be said to ‘reason’, and whether they can truly ‘learn’ in context. In this talk, I will suggest that approaches from cognitive science can provide useful tools for approaching these questions. Specifically, I will focus on rational analysis: analyzing behavior as a rational adaptation to an environment.

I’ll  first illustrate these ideas by discussing some of our work analyzing how the in-context learning abilities of language models can emerge as a rational adaptation to simple properties of the data distribution they are trained on. I’ll then discuss how this analysis extends to suggest that many of language models broader success and failures can be interpreted as rational responses to the natural data distribution — including cases where language models exhibit some of the same patterns of content entangled reasoning that humans do. Finally, I’ll illustrate how rational analysis implies that the contextual abilities of language models can overcome failures like the “reversal curse” — and how to exploit these insights to improve generalization.

Coming from Longwood? Sign up to take the shuttle.