How do you know if a method is better, or just has better hyperparameters? He He, Kyunghyun Cho, and I give a new tool to answer this in our #NAACL2024 paper: "Show Your Work with Confidence" arxiv.org/abs/2311.09480. Use it in your own work with just a "pip install opda"! 🧵 1/8

1/🚀 Introducing an identifiability theory for latent variable models that compare patterns across datasets. Can we identify salient variation that is specific to a data set, and under which conditions? With wonderful collaborators Jan-Christian Huetter, Ehsan Hajiramezanali, Jonathan Pritchard, and Aviv Regev

Our work "Protein Discovery with Discrete Walk-Jump Sampling" won an Outstanding Paper award ICLR 2025 !!! 🥳 Once again, thanks to my Prescient Design and Genentech colleagues, especially Dan Berenberg and Saeed Saremi.

Excited to share our new preprint 🥳🥳 Implicit guidance with PropEn: Match your data to follow the gradient 🔗 arxiv.org/pdf/2405.18075 Joint work with Andreas Loukas, Kyunghyun Cho, VGligorijevic at @prescientdesign & Genentech

Why does multi-modal modeling struggle compared to using a single modality or naive combinations of multiple modalities? Taro Makino, Sumit Chopra, Kyunghyun Cho, and I reveal factors behind these challenges and proposes a modality-agnostic framework to overcome them. 🧵1/7

📢 ICYMI: Our cofounder and Senior Director of Frontier Research Kyunghyun Cho's invited #ICLR2024 Keynote Talk on our "Lab-in-the-Loop for Antibody Design" is now available online for public viewing! 🎥 iclr.cc/virtual/2024/i…

AI molecule design systems are hard to test end-to-end bc experiments are slow and $$$. Approximate feedback from models and simulations is inaccurate and still too slow! In new work we propose closed-form test functions for bio sequence optimization arxiv.org/abs/2407.00236 1

Happy to announce that in September 2025, I will open my research laboratory at NYU Courant 🏢🔬. As an assistant professor of CS & Biology 💻🧬, I will carry on my work on advancing ML research for molecular biology, and apply those methods for scientific discoveries 🌟🔍.

delighted to have had the opportunity to speak with Julian Anna Nowogrodzki, nature about our recent initiative to develop our own LLMs in-house at Prescient Design and Genentech!

#ICML24 Training on AI-generated data destroys scaling laws; mixing of real & AI-data leads to transient training plateaus! Interested? Come to our poster Thu 1:30pm "Tale of Tails: Model Collapse as a Change of Scaling Laws" w Elvis Dohmatob Yunzhen Feng Pu Yang 杨 埔 François Charton

1/ How to evaluate automatically designed catalysts? Challenge: multiple catalysts work well but usually only one ground truth. CatScore: use a reverse prediction model to compare catalysts' predicted product to target product. w/ Kyunghyun Cho now out at Digital Discovery!

1/ 🧬 Excited to share CHEAP, our new work on compressed protein embeddings. We characterize the joint distribution of p(sequence, structure) in ESMFold's latent space, and find cool tidbits on compressibility, tokenizability, and pathologies: biorxiv.org/content/10.110… 🧵

it all started with the question: "how did they choose hyperparameters in (so-called) continual learning?" not even saying this is the right approach for model selection in continual learning. it's only "a" way to do so, but something needs to be done for model selection for

do we want to know which variables are direct causes of a target outcome, or the full dependencies among all variables? gradually i started to think that it's probably neither, since the utility of each cause is not a function of the distance to the target outcome variable but

I’d like to share a preprint on a topic close to my heart: learning to generalize ood. With Karolis Martinkus Ed Wagstaff Kyunghyun Cho We ask, what’s the worst-case performance of a model across any diverse test distribution within a domain? t.ly/BO2jQ

CDS & NYU Courant researchers Divyam Madaan, Taro Makino, Sumit Chopra, & Kyunghyun Cho introduced I2M2, a new framework that improves multi-modal AI performance across healthcare and vision-language tasks. nyudatascience.medium.com/new-framework-…

It’s time to use AI to help cancer patients get truly personalized treatment! To solve this, I founded Ataraxis AI with Krzysztof Geras. Ataraxis is an AI precision medicine company building a new era of tools for better treatment selection, starting with Ataraxis Breast:

EXCLUSIVE: Startup Ataraxis, which emerged from stealth Thursday, has developed a model that can predict the risk severity of breast cancer up to 30% more accurately than current tests. trib.al/Cur9UV6 trib.al/Cur9UV6

We (Jan Witowski and I) co-founded a startup! We want to personalize cancer treatment for every patient using AI. It's a fascinating research journey. I am very proud of what we achieved as a team. There is still so much to be done. Join us! jobs.ashbyhq.com/ataraxis-ai