Tal Linzen
@tallinzen.bsky.social
NYU professor, Google research scientist. Good at LaTeX.
Reposted by Tal Linzen
Really big announcement! See @wtimkey.bsky.social's thread for the details on an exciting new preprint from the NYU-UMass Syntactic Ambiguity Processing group. It is the culmination of the team's research efforts over these last couple of years, and we're really happy with it.
New Preprint: osf.io/eq2ra
Reading feels effortless, but it's actually quite complex under the hood. Most words are easy to process, but some words make us reread or linger. It turns out that LLMs can tell us about why, but only in certain cases... (1/n)
Reading feels effortless, but it's actually quite complex under the hood. Most words are easy to process, but some words make us reread or linger. It turns out that LLMs can tell us about why, but only in certain cases... (1/n)
November 14, 2025 at 7:53 PM
Really big announcement! See @wtimkey.bsky.social's thread for the details on an exciting new preprint from the NYU-UMass Syntactic Ambiguity Processing group. It is the culmination of the team's research efforts over these last couple of years, and we're really happy with it.
Reposted by Tal Linzen
New Preprint: osf.io/eq2ra
Reading feels effortless, but it's actually quite complex under the hood. Most words are easy to process, but some words make us reread or linger. It turns out that LLMs can tell us about why, but only in certain cases... (1/n)
Reading feels effortless, but it's actually quite complex under the hood. Most words are easy to process, but some words make us reread or linger. It turns out that LLMs can tell us about why, but only in certain cases... (1/n)
November 14, 2025 at 7:19 PM
New Preprint: osf.io/eq2ra
Reading feels effortless, but it's actually quite complex under the hood. Most words are easy to process, but some words make us reread or linger. It turns out that LLMs can tell us about why, but only in certain cases... (1/n)
Reading feels effortless, but it's actually quite complex under the hood. Most words are easy to process, but some words make us reread or linger. It turns out that LLMs can tell us about why, but only in certain cases... (1/n)
Reposted by Tal Linzen
New NeurIPS paper! Why do LMs represent concepts linearly? We focus on LMs's tendency to linearly separate true and false assertions, and provide an analysis of the truth circuit in a toy model. A joint work with Gilad Yehudai, @tallinzen.bsky.social, Joan Bruna and @albertobietti.bsky.social.
October 24, 2025 at 3:19 PM
New NeurIPS paper! Why do LMs represent concepts linearly? We focus on LMs's tendency to linearly separate true and false assertions, and provide an analysis of the truth circuit in a toy model. A joint work with Gilad Yehudai, @tallinzen.bsky.social, Joan Bruna and @albertobietti.bsky.social.
Reposted by Tal Linzen
🌍Introducing BabyBabelLM: A Multilingual Benchmark of Developmentally Plausible Training Data!
LLMs learn from vastly more data than humans ever experience. BabyLM challenges this paradigm by focusing on developmentally plausible data
We extend this effort to 45 new languages!
LLMs learn from vastly more data than humans ever experience. BabyLM challenges this paradigm by focusing on developmentally plausible data
We extend this effort to 45 new languages!
October 15, 2025 at 10:53 AM
🌍Introducing BabyBabelLM: A Multilingual Benchmark of Developmentally Plausible Training Data!
LLMs learn from vastly more data than humans ever experience. BabyLM challenges this paradigm by focusing on developmentally plausible data
We extend this effort to 45 new languages!
LLMs learn from vastly more data than humans ever experience. BabyLM challenges this paradigm by focusing on developmentally plausible data
We extend this effort to 45 new languages!
Reposted by Tal Linzen
Another banger from @tallinzen.bsky.social .
Also fits with some of the criticisms of Centaur and my faculty-based approach generally; if you want LLMs to model human cognition, give them more architecture akin to human faculty psychology like long and short-term memory.
arxiv.org/abs/2510.05141
Also fits with some of the criticisms of Centaur and my faculty-based approach generally; if you want LLMs to model human cognition, give them more architecture akin to human faculty psychology like long and short-term memory.
arxiv.org/abs/2510.05141
To model human linguistic prediction, make LLMs less superhuman
When people listen to or read a sentence, they actively make predictions about upcoming words: words that are less predictable are generally read more slowly than predictable ones. The success of larg...
arxiv.org
October 15, 2025 at 1:58 PM
Another banger from @tallinzen.bsky.social .
Also fits with some of the criticisms of Centaur and my faculty-based approach generally; if you want LLMs to model human cognition, give them more architecture akin to human faculty psychology like long and short-term memory.
arxiv.org/abs/2510.05141
Also fits with some of the criticisms of Centaur and my faculty-based approach generally; if you want LLMs to model human cognition, give them more architecture akin to human faculty psychology like long and short-term memory.
arxiv.org/abs/2510.05141
Reposted by Tal Linzen
Linguistics PhD student @jacksonpetty.org finds LLMs "quiet-quit" when instructions get long, switching from reasoning to guesswork.
With CDS' @tallinzen.bsky.social, @shauli.bsky.social, @lambdaviking.bsky.social, @michahu.bsky.social, and Wentao Wang.
nyudatascience.medium.com/llms-switch-...
With CDS' @tallinzen.bsky.social, @shauli.bsky.social, @lambdaviking.bsky.social, @michahu.bsky.social, and Wentao Wang.
nyudatascience.medium.com/llms-switch-...
LLMs Switch to Guesswork Once Instructions Get Long
LLMs abandon reasoning for guesswork when instructions get long, new work from Linguistics PhD student Jackson Petty & CDS shows.
nyudatascience.medium.com
September 10, 2025 at 3:26 PM
Linguistics PhD student @jacksonpetty.org finds LLMs "quiet-quit" when instructions get long, switching from reasoning to guesswork.
With CDS' @tallinzen.bsky.social, @shauli.bsky.social, @lambdaviking.bsky.social, @michahu.bsky.social, and Wentao Wang.
nyudatascience.medium.com/llms-switch-...
With CDS' @tallinzen.bsky.social, @shauli.bsky.social, @lambdaviking.bsky.social, @michahu.bsky.social, and Wentao Wang.
nyudatascience.medium.com/llms-switch-...
Reposted by Tal Linzen
DO NOT GIVE UP!
Our advocacy is working.
A key Senate committee has indicated that it will reject Trump’s proposed cuts to science agencies including NASA and the NSF.
Keep speaking up and calling your electeds 🗣️🗣️🗣️
Our advocacy is working.
A key Senate committee has indicated that it will reject Trump’s proposed cuts to science agencies including NASA and the NSF.
Keep speaking up and calling your electeds 🗣️🗣️🗣️
July 11, 2025 at 7:03 PM
DO NOT GIVE UP!
Our advocacy is working.
A key Senate committee has indicated that it will reject Trump’s proposed cuts to science agencies including NASA and the NSF.
Keep speaking up and calling your electeds 🗣️🗣️🗣️
Our advocacy is working.
A key Senate committee has indicated that it will reject Trump’s proposed cuts to science agencies including NASA and the NSF.
Keep speaking up and calling your electeds 🗣️🗣️🗣️
Congratulations to @linguistbrian.bsky.social for receiving this grant to study how to constrain language models to read complex sentences more like humans, and congratulations to me for getting to collaborate with him for another four years! www.umass.edu/humanities-a...
Brian Dillon Receives NSF Grant to Explore AI and Human Language Processing : College of Humanities & Fine Arts : UMass Amherst
Linguist Brian Dillon receives NSF grant to investigate how AI and humans differ in interpreting meaning during language comprehension.
www.umass.edu
July 11, 2025 at 9:30 PM
Congratulations to @linguistbrian.bsky.social for receiving this grant to study how to constrain language models to read complex sentences more like humans, and congratulations to me for getting to collaborate with him for another four years! www.umass.edu/humanities-a...
My Twitter account has been hacked :( Please don't click on any links "I" posted on that account recently!
July 2, 2025 at 4:20 PM
My Twitter account has been hacked :( Please don't click on any links "I" posted on that account recently!
I'm hiring at least one post-doc! We're interested in creating language models that process language more like humans than mainstream LLMs do, through architectural modifications and interpretability-style steering. Express interest here: docs.google.com/forms/d/e/1F...
NYU LLM + cognitive science post-doc interest form
Tal Linzen's group at NYU is hiring a post-doc! We're interested in creating language models that process language more like humans than mainstream LLMs do, through architectural modifications and int...
docs.google.com
June 21, 2025 at 3:13 PM
I'm hiring at least one post-doc! We're interested in creating language models that process language more like humans than mainstream LLMs do, through architectural modifications and interpretability-style steering. Express interest here: docs.google.com/forms/d/e/1F...
Reposted by Tal Linzen
How well can LLMs understand tasks with complex sets of instructions? We investigate through the lens of RELIC: REcognizing (formal) Languages In-Context, finding a significant overhang between what LLMs are able to do theoretically and how well they put this into practice.
June 9, 2025 at 6:02 PM
How well can LLMs understand tasks with complex sets of instructions? We investigate through the lens of RELIC: REcognizing (formal) Languages In-Context, finding a significant overhang between what LLMs are able to do theoretically and how well they put this into practice.
Reposted by Tal Linzen
Following the success story of BabyBERTa, I & many other NLPers have turned to language acquisition for inspiration. In this new paper we show that using Child-Directed Language as training data is unfortunately *not* beneficial for syntax learning, at least not in the traditional LM training regime
“Child-Directed Language Does Not Consistently Boost Syntax Learning in Language Models”
I’m happy to share that the preprint of my first PhD project is now online!
🎊 Paper: arxiv.org/abs/2505.23689
I’m happy to share that the preprint of my first PhD project is now online!
🎊 Paper: arxiv.org/abs/2505.23689
Child-Directed Language Does Not Consistently Boost Syntax Learning in Language Models
Seminal work by Huebner et al. (2021) showed that language models (LMs) trained on English Child-Directed Language (CDL) can reach similar syntactic abilities as LMs trained on much larger amounts of ...
arxiv.org
May 30, 2025 at 8:45 PM
Following the success story of BabyBERTa, I & many other NLPers have turned to language acquisition for inspiration. In this new paper we show that using Child-Directed Language as training data is unfortunately *not* beneficial for syntax learning, at least not in the traditional LM training regime
Cross-posting the abstracts for two talks I'm giving next week! This one on formal languages for LLM pretraining and evaluation, at Apple ML Research in Copenhagen on Wednesday
May 23, 2025 at 1:44 PM
Cross-posting the abstracts for two talks I'm giving next week! This one on formal languages for LLM pretraining and evaluation, at Apple ML Research in Copenhagen on Wednesday
Updated version of our position piece on how language models can help us understand how people learn and process language, on why it's crucial to train models on cognitive plausible datasets, and on the BabyLM project that addresses this issue.
📣Paper Update 📣It’s bigger! It’s better! Even if the language models aren’t. 🤖New version of “Bigger is not always Better: The importance of human-scale language modeling for psycholinguistics” osf.io/preprints/ps...
OSF
osf.io
May 12, 2025 at 4:14 PM
Updated version of our position piece on how language models can help us understand how people learn and process language, on why it's crucial to train models on cognitive plausible datasets, and on the BabyLM project that addresses this issue.
Reposted by Tal Linzen
At #HSP2025, I'll present work with @tallinzen.bsky.social and @shravanvasishth.bsky.social on modeling garden-pathing in a huge benchmark dataset: hsp2025.github.io/abstracts/29.... Statistically decomposing the effect into subprocesses greatly improves predictive fit over just comparing means!
hsp2025.github.io
March 14, 2025 at 10:17 AM
At #HSP2025, I'll present work with @tallinzen.bsky.social and @shravanvasishth.bsky.social on modeling garden-pathing in a huge benchmark dataset: hsp2025.github.io/abstracts/29.... Statistically decomposing the effect into subprocesses greatly improves predictive fit over just comparing means!
Going to give this website another shot! What are good lists of linguistics, psycholinguistics, NLP and AI accounts?
March 27, 2025 at 7:14 PM
Going to give this website another shot! What are good lists of linguistics, psycholinguistics, NLP and AI accounts?
Thanks Ted for mentioning me in the same tweet as Chris! This website really is better than the other one!
Another good sign for this platform is that people like @tallinzen.bsky.social and @chrmanning.bsky.social are dipping their toes in it (even if no pfps yet). Give them a follow if you know who they are.
November 19, 2023 at 5:52 AM
Thanks Ted for mentioning me in the same tweet as Chris! This website really is better than the other one!
Very little happening on here but silence is certainly better than all of the boardroom drama takes on the other website. Four different people I follow just came up with the same unfunny joke about the most recent development in the drama, apparently independently?
November 19, 2023 at 5:46 AM
Very little happening on here but silence is certainly better than all of the boardroom drama takes on the other website. Four different people I follow just came up with the same unfunny joke about the most recent development in the drama, apparently independently?
Reposted by Tal Linzen
Do deep transformer LMs generalize better? In a new preprint we (Sjoerd van Steenkiste, Ishita Dasgupta, Fei Sha, Dan Garrette, & @tallinzen.bsky.social) control for parameter count to show how depth helps models on compositional generalization tasks, but diminishingly so 🧵
jacksonpetty.org/depth
jacksonpetty.org/depth
The Impact of Depth and Width on Transformer Language Model Generalization
To process novel sentences, language models (LMs) must generalize compositionally -- combine familiar elements in new ways. What aspects of a model's structure promote compositional generalization? Fo...
jacksonpetty.org
November 10, 2023 at 5:05 PM
Do deep transformer LMs generalize better? In a new preprint we (Sjoerd van Steenkiste, Ishita Dasgupta, Fei Sha, Dan Garrette, & @tallinzen.bsky.social) control for parameter count to show how depth helps models on compositional generalization tasks, but diminishingly so 🧵
jacksonpetty.org/depth
jacksonpetty.org/depth