Tal Linzen

LLMs run on vectors, not symbols. A preprint led by Yale’s Tom McCoy (@rtommccoy.bsky.social) with CDS Assoc. Prof. @tallinzen.bsky.social found those vectors closely approximate symbolic structure in language, arithmetic, logic, and code. arxiv.org/abs/2608.29530 news.yale.ed…

The Emergent Symbolic Structure of Artificial Neural Networks arxiv.org

AI Weekly's analysis →

• A new preprint reports that neural networks' vector representations can be closely approximated by closed-form symbolic equations with behavior 'largely unchanged.'
• The substitution was tested on small networks trained to manipulate lists and on LLMs across arithmetic, logic, computer code, and language.
• The authors say the symbolic approximation also lets them steer an LLM's behavior via targeted interventions on its internal representations.

Read full analysis →

View on Bluesky →

LLMs run on vectors, not symbols. A preprint led by Yale’s Tom McCoy (@rtommccoy.bsky.social) with CDS Assoc. Prof. @tallinzen.bsky.social found those vectors closely approximate symbolic structure in language, arithmetic, logic, and code. arxiv.org/abs/2608.29530 news.yale.ed…

How do AI chatbots ‘reason’ like humans? A Yale-led study offers clues news.yale.edu

AI Weekly's analysis →

Read full analysis →

View on Bluesky →

Updated preprint with @tallinzen.bsky.social and @shravanvasishth.bsky.social : "LLM surprisal is necessary but not sufficient to capture English garden-path effects: Evidence from joint latent modeling of reading paradigms" arxiv.org/abs/2602.04489

Deconstructing sentence disambiguation by joint latent modeling of reading paradigms: LLM surprisal is not enough arxiv.org

AI Weekly's analysis →

Read full analysis →

View on Bluesky →

New paper! LLM memory keeps improving, but this makes them *worse* as user sims. If we want to build models that can, e.g., simulate realistic students to train chatbots to be better teachers, then these models need to be able to forget like humans do 📄: arxiv.org/abs/2605.25680

[2605.25680] Simulating Human Memory with Language Models arxiv.org View on Bluesky →

Transformer LMs develop attention heads that appear to track syntactic dependencies. But do these heads actually drive syntax-sensitive predictions? In a new preprint with @tallinzen.bsky.social, we put this to the test. The answer was surprisingly mixed. 🧵 Paper link: arxiv.o…

Do Syntax Heads Influence Model Behavior? arxiv.org View on Bluesky →

Can LLMs introspect? Anthropic said yes. However, CDS PhD student Shashwat Singh, CDS Associate Professor Tal Linzen (@tallinzen.bsky.social) & CDS Faculty Fellow Shauli Ravfogel (@shauli.bsky.social) found the evidence falls short. nyudatascience.medium.com/cds-research...

nyudatascience.medium.com View on Bluesky →

To make AI better at imitating humans, limit its memory. Ex-CDS Faculty Fellow Nicholas Tomlin, CDS PhD @michahu.bsky.social, & CDS Prof @tallinzen.bsky.social show that capping a model's memory at 4 slots makes it a far better stand-in for a real person. nyudatascience.medium…

nyudatascience.medium.com View on Bluesky →