New paper! The Linear Representation Hypothesis is a powerful intuition for how language models work, but lacks formalization. We give a mathematical framework in which we can ask and answer a basic question: how many features can be stored under the hypothesis? 🧵 arxiv.org/abs/2602.11246
Excited to share a new working paper!
What happened when Change.org integrated an AI writing tool into their platform? We provide causal evidence that petition text changed significantly while outcomes did not improve. 1/
arxiv.org/abs/2511.13949
another banger from @louisathomas.bsky.social
www.newyorker.com/sports/sport...
New #NeurIPS2025 paper: how should we evaluate machine learning models without a large, labeled dataset? We introduce Semi-Supervised Model Evaluation (SSME), which uses labeled and unlabeled data to estimate performance! We find SSME is far more accurate than standard methods.
I am on the job market this year! My research advances methods for reliable machine learning from real-world data, with a focus on healthcare. Happy to chat if this is of interest to you or your department/team.
I've been working for many months on this article on Silicon Valley's under-the-radar role in bringing AI into schools across the US. I really hope you'll read it — here's a gift link — but I'll tell you some of the highlights in this thread. (1/x)
🚨 New postdoc position in our lab at Berkeley EECS! 🚨
(please reshare)
We seek applicants with experience in language modeling who are excited about high-impact applications in the health and social sciences!
More info in thread
1/3
What a crossover!
This is great, & there's clear analogy to the burgeoning mechanism design community for AI alignment: who is providing RLHF votes? Do their preferences reflect yours? Discussions about social choice and collective constitutions are interesting, but "what and who is in the data" is just as important.
This is amazing
They're in their move fast and break things era 🙃
This take emerged organically from just how well our method on SAEs for hypothesis generation (HypotheSAEs) performed, which surprised all of us!
See the paper arxiv.org/abs/2506.23845
Thanks @kennypeng.bsky.social, Jon, @emmapierson.bsky.social, @nkgarg.bsky.social for another nice collaboration.
This capability of discovering unknown concepts opens many opportunities for applied machine learning. We can design better whitebox predictors, better audit high-stakes models for bias, and generate hypotheses for CSS research. More broadly, SAEs can help bridge the "prediction-explanation" gap.
These tasks lie in contrast to probing, where we're trying to predict the presence of a *known* concept; and steering, where we're trying to include a *known* concept in an LLM output. SAEs lose to simple baselines on these tasks. (2 good papers on this: "AxBench" and Kantamneni, Engels et al. 2025)
How do we reconcile our view with recent negative results? Our key distinction is that SAEs are useful when you don't know what you're looking for: how does my text classifier predict which headlines will go viral? How does my LLM perform addition? These are "unknown unknowns".
📢New POSITION PAPER: Use Sparse Autoencoders to Discover Unknown Concepts, Not to Act on Known Concepts
Despite recent results, SAEs aren't dead! They can still be useful to mech interp, and also much more broadly: across FAccT, computational social science, and ML4H. 🧵
Nice work! Cool to see that item difficulty predicts human-llm disagreement. We also studied similar questions with the DICES dataset: aclanthology.org/2024.emnlp-m...
Heat map showing that more accurate models have more correlated errors.
Are LLMs correlated when they make mistakes? In our new ICML paper, we answer this question using responses of >350 LLMs. We find substantial correlation. On one dataset, LLMs agree on the wrong answer ~2x more than they would at random. 🧵(1/7)
arxiv.org/abs/2506.07962
@jessica.bsky.social on individual reporting as a means to build collective knowledge.
ARR question: If I submit to a cycle, how long do those reviews "last"? e.g. if I submit to the July cycle but can't go to AACL, can I commit my July reviews to the conference associated with the next (October) cycle? @aclrollingreview.bsky.social
A gif explaining the value of test-time augmentation to conformal classification. The video begins with an illustration of TTA reducing the size of the predicted set of classes for a dog image, and goes on to explain that this is because TTA promotes the true class's predicted probability to be higher, even when it's predicted to be unlikely.
New work 🎉: conformal classifiers return sets of classes for each example, with a probabilistic guarantee the true class is included. But these sets can be too large to be useful.
In our #CVPR2025 paper, we propose a method to make them more compact without sacrificing coverage.
I would like to spend up to 5-10 hours to learn about basic macroeconomics (I know it's maybe fake, but setting that aside for a moment...). Does anyone have any recommendations?
Huge congrats, Marianne!!
I find that I've actually gone out of my way to stop using bullet points in reviews now because Any Review With Bullet Points is a Bot 🥲
People love to hate on the transition 3-pointer as evidence of how the 3 has ruined basketball, but I think it's usually just the right play... if you have numbers in transition, your teammate can easily get a putback off a miss, so might as well try the 3
We'll present HypotheSAEs at ICML this summer! 🎉
Draft: arxiv.org/abs/2502.04382
We're continuing to cook up new updates for our Python package: github.com/rmovva/Hypot...
(Recently, "Matryoshka SAEs", which help extract coarse and granular concepts without as much hyperparameter fiddling.)
So awesome, congrats Lucy!!! 🧀
Yesterday's Game 6 was depressing, and this article precisely delineated the reasons why. And sometimes, a precise retelling of what you're feeling is all you need to feel better. www.nytimes.com/athletic/633... @thompsonscribe.bsky.social
Did you take the hot air balloon pic?!
Check out Erica's nice work. They not only develop a well-grounded model for disparities in disease progression, but also conduct experiments with real NYP cardiology data! (Anyone who works in healthcare knows how much of a feat it is to use data other than MIMIC)
