Dear colleagues, I'm recruiting at least one post-doc for a project at New York University aimed at creating language models that process language more like humans than mainstream LLMs do <https://tallinzen.net/media/papers/huang_et_al_2024_jml.pdf>. We are planning to explore architectural modifications, training data interventions, and steering through interpretability.
One motivation for this project is the empirical finding <https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00548/115371/Why-Does-Surprisal-From-Larger-Transformer-Based> that the better LLMs become in terms of perplexity and task performance, the worse they are as cognitive models of how people read and learn language; we think that to reverse this trend we need to find ways to constrain them (in terms of e.g. working memory, parse parallelism, and factual and linguistic knowledge), and improve them in other ways to make up for these constraints, e.g. through increasing data efficiency <https://tallinzen.net/media/papers/wilcox_et_al_2025_jml.pdf>. We're planning to benchmark the models against behavioral and neural data from humans: eyetracking, fMRI and intracranial recordings. Some of the data already exists, and some will be collected by collaborators at other universities specifically for this project. But we also expect to do a lot of fundamental modeling and interpretability work. You do not need to have existing experience in cognitive science, but you should have a strong track record in computational research; and you should be interested in using AI for science, in learning about cognitive science and collaborating with linguistics and cognitive scientists, and in doing open-ended fundamental research on LLMs. There are no teaching requirements. The position will be renewed every year, but we expect the funding for this project to last four years. You will be affiliated with NYU's Center for Data Science, and, if relevant, also with the department of linguistics. NYU has large NLP and computational cognitive science communities, with lots of opportunities for collaborations. The start date is flexible, though of course you should have a PhD by the time you start. Your application is most likely to be considered if you apply before *August 10th.* Please fill out this lightweight form <https://docs.google.com/forms/d/e/1FAIpQLSc5IwTU43CWVjQYsWbvPkDFH7dFKglqRfPdWRJSvCbYuxlv-A/viewform> to express interest, and you can also email me directly if they have any questions. I'll be at ACL 2025 and am happy to chat about the position. If you're interested in working together but don't exactly fit the description, don't hesitate to reach out! -- Tal Linzen <https://tallinzen.net/> Associate Professor of Linguistics and Data Science New York University
_______________________________________________ Corpora mailing list -- corpora@list.elra.info https://list.elra.info/mailman3/postorius/lists/corpora.list.elra.info/ To unsubscribe send an email to corpora-le...@list.elra.info
