Tal Linzen
New York University, NY·Computer Science·Early-career
People understand sentences they have never seen before, and they do it with limited memory and far less practice than today's AI systems need. AI language models such as ChatGPT handle language well, yet they often process it differently from people.
- Fields
Computer Science
Psychology
Neuroscience
- Topics
computational linguistics
psycholinguistics
language models
- Grants
- 3 known grants, through Aug 2029
What they work on
- Explain why language models that predict the next word better match human reading times worse, by comparing model predictability estimates with human reading data.
- Test whether training goals based on meaning and causal interventions, which edit a model's internal states, can make language models keep only a few interpretations of an ambiguous sentence, as humans seem to.
- Run psycholinguistic experiments with new temporarily ambiguous sentences to test whether the modified models predict where human readers struggle.
- Build transformer language models with human-like working memory limits, and compare them with people in memory-retrieval and sentence-processing experiments.
- Test whether memory limits help models pick up language structure, by comparing the models' learning path and efficiency with children's language acquisition.
- Record fMRI and intracranial brain activity while people process sentences that link distant words, and test how well memory-limited models explain these data and existing EEG and MEG data.
- Create semantic parsing tasks in many languages, including linguist-designed grammars paired with lambda calculus meanings, with training and test sets that differ in sentence structure.
- Use meta-learning on these tasks to give neural networks a preference for structural generalization, and examine how initial weights, training dynamics and loss landscape produce it.
More about the research
This research asks where the differences lie, and how language models could be built to process language more like humans do. Closer models could help explain how the mind and brain handle language, and they could also make AI learn more efficiently and generalize more reliably. Three projects share this goal. One looks at how readers settle on a meaning, one looks at working memory, and one looks at how systems combine word meanings into sentence meanings.
Why it matters
Language models that process language the way people do could help scientists understand how the brain handles language. The grants say this could eventually inform treatments for language impairments. More human-like models may also need less training data. Systems that reliably interpret complex, unfamiliar sentences could make virtual assistants and other language technologies work better across many languages, including those spoken by smaller communities.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Collaborative Research: Research: CompCog: RI: Medium: Semantic Focusing: Controlling LM Interpretations for Human-Model Alignment
NSF · Sep 2025 – Aug 2029
CRCNS: Developing and Testing Language Models with Cognitively Plausible Memory
NIH · Aug 2025 – Jul 2029
CAREER: RI: Structural Linguistic Generalization Through Expert-Designed Tasks
NSF · Sep 2023 – Aug 2028
Closest labs
Brian Dillon
University of Massachusetts, MA·Psychology
Explain why a language model's perplexity, its skill at predicting the next word, correlates positively with worse fit to human reading times, by comparing model predictability estimates against reading-time data.
Satoshi Tomioka
University of Delaware, DE·Social Science
Test whether the difficulty of alternative word orders comes from filler-gap dependency formation, the effort of linking a displaced phrase back to the position it moved from, using self-paced reading across several sentence types.
Kyle Mahowald
University of Texas at Austin, TX·Social Science
Measure how speakers of a variety of languages choose deictic words such as "this," "that," "here," and "there" to refer to objects and locations, using cross-linguistic behavioral experiments.
The school
New York University: 72 labs on PI Pages, $1.5B a year on research, 540 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.