Jacob Andreas
I'm interested in language as a tool for communication and reasoning. Humans language from remarkably little data. Once learned, language is the most distinctive and powerful feature of human cognition: we use it to acquire new information, coordinate with others, and structure our thoughts. Over the last decade, language has also become the primary vehicle by which AI systems learn and reason. My lab builds tools for understanding how LMs work, develops techniques for specifying and steering their behavior, and tries to figure out what they can teach us about language learning and understanding in humans.
I'm an associate professor at MIT in EECS and CSAIL. I did my PhD work at Berkeley, where I was a member of the Berkeley NLP Group and the Berkeley AI Research Lab. I've also spent time with the Cambridge NLIP Group, and the NLP Group and the (erstwhile) Center for Computational Learning Systems at Columbia.
Prospective students and visitors: please see the contact page below. Read my advising statement if you're considering applying!
Curriculum vitæ, Google scholar, accessibility @ MIT
Lab / Research / Bio / Teaching / Contact
Some representative research directions:
Understanding neural models
My lab is probably best known for work on deep network interpretability. We developed the first "automated interpretability" techniques, which train neural nets to translate neural features and representations into natural language [ADK18, HSB+22]. We also provided some of the early characterizations of in-context learning [ASA+22] and world modeling capabilities in LMs [LNA21, LGA25]. These days we're interested in automated interpretability techniques that explain models using code rather than language [HLA26], and in "introspective" approaches to interpretability [LGH+26].
Specification and monitoring
Current deep network models are expressive enough to do more or less whatever we want when trained on enough of the right kind of data. Thus the primary challenges are (1) figuring out how to specify what these models should actually do, and (2) verifying that they do it. Much of our recent work focuses on interactive preference elicitation methods that can learn richer and more reliable models of human preferences than people can produce unassisted [L*T*GA25, GPTA26] and methods that enable models to accurately self-report behavioral tendencies [D*P*S+26, PRG+26].
Collaboration graph trivia: My Erdős number is at most three (J Andreas to R Kleinberg to L Lovász to P Erdős). My Kevin Bacon number (and consequently my Erdős-Bacon number) remains lamentably undefined, but my Kevin Knight number (it's a thing) is one. I have never starred in a film with Kevin Knight. Noam Chomsky is my great-great-grand-advisor (J Andreas to D Klein to C Manning to J Bresnan to N Chomsky).
Photo: Gretchen Ertl / MIT CSAIL