Angelica Chen

New work w/@sadhikamalladi, @lilyhzhang, @xinyichen2, @QiuyiRichardZ, Rajesh Ranganath, @kchonyc: Contrary to conventional wisdom, RLHF/DPO does *not* produce policies that mostly assign higher likelihood to preferred responses than to less preferred ones.

Generalist AIs ace medical exams but fail at hospital operations. But a small, specialized model created by @lavenderjiang99, @_angie_chen, @kchonyc, @ekoermann, et al beat massive generalists in clinical settings. nyudatascience.medium.com/generalist-ais…

@kchonyc @joshwoodward

please @_angie_chen fix it natively please

Hiring alert 🚨 If you’re working on LLMs—especially reasoning, RL, or distillation—and want to collaborate on some exciting research, we’d love to hear from you. Apply below.

We built Lang1: a 100M–7B family of models specialized for hospital operations. After finetuning, Lang1-1B outperforms generalist models up to 671B, transfers to unseen tasks and another hospital, and is more data-efficient. 🧵Why small specialists win in healthcare.

🚨Prompt Curriculum Learning (PCL) - Efficient LLM RL training algo! - We investigate factors that affect convergence: bsz, # prompt, # gen, prompt selection - We propose PCL: lightweight algo that *dynamically selects intermediate-difficulty prompts* using a learned value model

@kchonyc @orf_bnw 😨

is gemini down? noooooooooooo, @orf_bnw @_angie_chen

Naomi Saphra is tracing AI outputs back to their origins. In a conversation with @benbenbrubaker, the Harvard researcher explains why an evolutionary lens can help researchers study interpretability. quantamagazine.org/to-understand-…

Working with Sadhika was one of the biggest highlights of my PhD and I couldn't be more excited to see her mentor the next generation of brilliant scientists! 🌟💫 Apply to work with her at UCSD!

@kchonyc @orf_bnw 🤔 @vqctran @YiTayML maybe this should be fixed

great job, @orf_bnw and @_angie_chen !

it is amazing to have @tallinzen as a colleague as he inspires so many people including myself and gemini 2.5 pro. here's the new position paper on Cardiology of Language Models written almost entirely by Gemini 2.5 Pro with my promp ... nope ... supervision. (and thanks to @gregd_nlp for motivation!)

@infoxiao One of these sounds so familiar 🤔

I've been automating parts of my research workflow. Sometimes amazing, sometimes I'm the engineer spending 3 hours to automate a 30-minute task. I think there might be a useful framework here: Context × Action ✅ Potential Sweet Spot (Low Context + High Action): "Run 10 evals across 20 checkpoints" → Clear task, lots of repetitive work ❌ Where It Seems to Break (High Context + Low Action): "Book me a flight but I prefer aisle seats, have Delta status, need to coordinate with friends..." Takes longer to explain than to just book it yourself. Maybe this explains why so many AI features feel useless? The magic might happen when explanation is minimal but the task list is long.

@vqctran @agihippo Happy birthday!!! 🎁🎊🎉 😄

@agihippo thank you thank you!!

Happy birthday @vqctran

LLMs vs Protein Design Tools: Who Wins? if you missed them @icmlconf, here's the poster for work led by @_angie_chen & @samuel_stanton_ that answers this question! Link to paper below 👇

Starting in 30 min at 10am PT! @nsaphra presents at DLCT: Sudden Drops in the Loss: Syntax Acquisition, Phase Transitions, and Simplicity Bias in MLMs Angelica Chen, Ravid Shwartz-Ziv, Kyunghyun Cho, Matthew L. Leavitt, Naomi Saphra arxiv.org/abs/2309.07311 Zoom 👇

Paper: arxiv.org/abs/2410.22296

We propose a new synthetic benchmark for highly-constrained biophysical optimization tasks, and propose a new semi-online RL algorithm for training an LLM to iteratively improve its own generations while using very few labels. Come chat with me and @samuel_stanton_ !

Have you ever wondered how your specialized biomolecule engineering model compares with a general purpose LLM on bio-optimization tasks? Come check out our work at ICML, in the East poster session (#E-2804) happening right now!. #ICML2025

@infoxiao @YiTayML @kchonyc @orf_bnw @OfficialLoganK 🥲 so much for expert choice routing

@YiTayML @kchonyc @_angie_chen @orf_bnw @OfficialLoganK Can’t be activated. Dead expert

i'm getting dizzy waiting for `gemma3n:4b-it-qat`. please, @_angie_chen

@infoxiao @kchonyc @orf_bnw the real purpose of PhD students

@kchonyc @orf_bnw Don't you have @_angie_chen now? Plx fix thanks.

Gemini team, can you please fix your rendering to remove "[cite_start]" everywhere? i need to copy-paste the result on a new chat to have Gemini clean up these unnecessary and wrong markers. @orf_bnw get to work, please

🌉 Bridging Offline & Online RL for LLMs 🌉 📝: arxiv.org/abs/2506.21495 New paper shows on verifiable & non-verifiable tasks: - Online DPO & GRPO give similar performance. - Semi-online (iterative) DPO with sync every s steps (more efficient!) works very well also. - Offline DPO is way behind. - Combining verifiable + non-verifiable works! Cross-transfer gains. - Recipes for how to make this work. 🧵1/4

Bridging Offline and Online Reinforcement Learning for LLMs Investigates the effectiveness of RL for finetuning LLMs when transitioning from offline to semi-online to fully online regimes for both verifiable and nonverifiable tasks.