Raj Movva

💡New preprint & Python package: We use sparse autoencoders to generate hypotheses from large text datasets. Our method, HypotheSAEs, produces interpretable text features that predict a target variable, e.g. features in news headlines that predict engagement. 🧵1/

Atul Butte died yesterday. The world lost a giant. A big bear of a man. With a huge smile. With love for everyone. With energy that could power a room. I loved everything about Atul. I loved how he was always happy. I loved how excited he was about science and helping people.

i wake up. something’s wrong with the clock on the wall. the numbers are jumbled. my hands aren’t right. i tell my wife. she responds: “that’s not just an observation—it’s a powerful insight.” i scream.

I've resolved this positively: 2 papers convincingly show sparse autoencoders beating baselines on real tasks: Hypothesis Generation & Auditing LLMs SAEs shine when you don't know what you're looking for, but lack precision. Sometimes the right tool for the job, sometimes not.

We're presenting two papers Wednesday at #ICML2025, both at 11am. Come chat about "Sparse Autoencoders for Hypothesis Generation" (west-421), and "Correlated Errors in LLMs" (east-1102)! Short thread ⬇️

1. We will present HypotheSAEs at #ICML2025, Wednesday 11am (West Hall B2-B3 #W-421). 2. Let me know if you'd like to chat about: - AI for hypothesis generation - why SAEs are still useful - whether PhD students should stay in school

This is a nice experiment: if you finetune an Othello next-move-predictor to reconstruct the board from its internal state, the reconstructed boards are often incorrect, but they have the same next moves as the true board! So next token prediction might be "too easy", in that a…

The reaction to this result shouldn't just be "what about Opus 4 / o5-ultrapro etc". For example, one takeaway is that human-AI collaboration for coding (what we want) isn't aligned with task-completion benchmarks (what we measure), and we should try to understand why!

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)

individual reporting for post-deployment evals — a little manifesto (& new preprints!) tldr: end users have unique insights about how deployed systems are failing; we should figure out how to translate their experiences into formal evaluations of those systems.

Cool exploration of how pretraining data shapes LLMs on medical topics. One result: clinical jargon doesn't show up much in pretraining, but is prevalent in clinical notes (a much-discussed use case for LLMs). Nice use of @allen_ai's What's In My Big Data tool.

Today is a good day to share my favorite KD fact, which is that he invested in @huggingface all the way back in... 2017. A true champion of open source AI 👑

Saddened by @atulbutte's passing. When I was in high school, he was gracious enough to do an interview for our student-run science magazine. There are many good reasons he could've said no, but he chose to spend his time inspiring teenagers. It worked on me. RIP, Prof. Butte.

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…