This is collectively joint work with @rajmovva.bsky.social, @nkgarg.bsky.social, Jon Kleinberg @emmapierson.bsky.social, Elliot Kim, and Avi Garg.

Come chat with us!

In Correlated Errors in Large Language Models, we show that LLMs are correlated in how they make mistakes. On one dataset, LLMs make the same mistake 2x more than random chance.

In Sparse Autoencoders for Hypothesis Generation, we show that SAEs can be used to find predictive natural language concepts in text data (e.g., that "addresses collective human responsibility" predicts lower headline engagement), achieving SOTA performance and efficiency.

This is joint work with Elliot Kim and Avi Garg (co-leads), and @nkgarg.bsky.social. We’ll be presenting at ICML in two weeks! Come see us at East Exhibition Hall A-B #E-2905 (11am-1:30pm Wednesday). (7/7)

The empirics here also add nuance to theory. While less correlated models can be more accurate together through a “wisdom of crowds,” this effect doesn't hold when newer, more correlated models are adopted in our simulations: gains from individual accuracy outweigh losses from homogeneity. (6/7)

However, in equilibrium, increased correlation—as predicted by past theoretical work—actually improves average applicant outcomes (intuitively, of the applicants who receive a job offer, more correlation means they will get more offers).

For the theory, see arxiv.org/abs/2312.09841 (5/7)

Since LLMs are correlated, this also leads to greater systemic exclusion in a labor market setting: more applicants are screened out of all jobs. Systemic exclusion persists even when different LLMs are used across firms. (4/7)

A consequence of error correlation is that LLM judges inflate accuracy of models less accurate than it. Here, we plot accuracy inflation against true model accuracy. Models from the same company (in red) are especially inflated. (3/7)

What explains error correlation? We found that models from the same company are more correlated. Strikingly, more accurate models also have more correlated errors, suggesting some level of convergence among newer models. (2/7)

This was lots of fun. Dogathoners: @gsagostini.bsky.social @sidhikabalachandar.bsky.social @ericachiang.bsky.social @nkgarg.bsky.social @laufer.bsky.social @rajmovva.bsky.social @emmapierson.bsky.social @shuvoms.bsky.social @dmshanmugam.bsky.social

Data: data.cityofnewyork.us/Health/NYC-D...

We also made a web app (github.com/shuvom-s/nyc...) you can run locally to make maps by name/breed.

We also briefly explored a “dog park simulator.”

7) Dogs are licensed the most in July and the least in November.

6) Dog names are shorter than baby names; 30% of dogs have names less than 5 letters long, but only 20% of babies.

5) Goldendoodles, Poodle Crossbreeds, and French Bulldogs are becoming more popular. Chihuahua's are becoming less popular.

Luna is the only popular name that is becoming more common among dogs AND babies.