The CDS Monthly Research Feature |
When a sentence misleads a reader, the eyes give it away: they stop and jump back to reread an earlier word. A new study led by William Timkey, a PhD student in the Department of Linguistics at New York University, shows that this backtracking is exactly the part of reading that today’s language models fail to predict, even as they capture a reader’s first, forward pass through the same sentence.
The paper, “Eye Movements Reveal a Dissociation between Prediction and Structural Processing Difficulty in Language Comprehension,” takes up a long-running question about what makes some sentences hard. Timkey wrote the paper with his advisor, CDS Associate Professor of Linguistics and Data Science Tal Linzen and others.
|
|
|
Some hiring managers no longer read the resumes they receive. They feed the entire applicant pool into ChatGPT and let it decide who advances. When CDS Assistant Professor of Computer Science, Engineering, and Data Science Emily Black first heard small business owners describe doing this, she was shocked, as this use was not at all represented in the academic literature. So Black joined forces with NYU Tandon Assistant Professor of Computer Science and Engineering Rosanna Bellini and NYU Tandon PhD student Sajel Surati, who she and Bellini co-advise, and the team set out to document what recruiters were actually doing.
|
A doctor who settles on a diagnosis early tends to order the tests that confirm the hunch rather than the ones that might rule it out, and the true cause can slip past unnoticed. The habit of gathering evidence to support a belief instead of trying to break it, named confirmation bias in a 1960 psychology study, turns out to shape how large language models reason as well. That is the finding of “Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models,” a paper led by recent NYU Courant computer science master’s alum Ayush Rajesh Jhaveri.
|
The genes that encode the cell’s ribosomes are among the most heavily used stretches of DNA in the human genome, copied hundreds of times across five chromosomes, and yet almost no disease has ever been pinned to a mutation in any of them. For a gene family this central, the silence in the medical literature was strange, since ribosomal defects such as those affecting mitochondrial ribosomes and ribosomal protein dosage are well known to cause illness. A study from the lab of NYU Professor of Biology Andreas Hochwagen argued that the mutations are there, sitting below the resolution of the tools used to find them. The work began when Hochwagen connected with three then-master’s students at CDS, Xufan Ma, Fiona Chow, and Buz Galbraith, through the CDS Spring Research Project Fair.
|
|
|
Copyright (C) 2026 Center for Data Science. All rights reserved.
|
|
|
|
