Big congratulations to Avi Wigderson of IAS Princeton for winning the Turing Award in CS. Truly an all-time great in theoretical computer science and discrete math. Also one of the nicest human beings I know --friend and mentor to so many (including me) tinyurl.com/fz5vxxaf

Our LabelBench work has been accepted to the DMLR journal🎉 Super smooth experience and highly recommended for anyone working on data centric ML work.

Check out labeltrain.ai: our broader set of label efficient learning work
More on LabelBench: x.com/jifan_zhang/st…

🆕 The Lancet
Counterfactual #AI models have exciting potential (and challenges) in medicine and life science
Why? An explainer, by Su-In Lee and me
thelancet.com/journals/lance…

A TWO-PLAYER system for online RL fine-tuning.

Honored to become a 2024 #SloanFellow . Thanks to Sloan Foundation and to all my students and collaborators for their amazing work!

“The problem with large language models is that they’re large!” In this UW-IT News story, #UWAllen ’s Tim Dettmers explains QLora, the Madrona prize-winning tool from UW NLP researchers that enables you to take an LLM and 'make it your own”: itconnect.uw.edu/making-languag… 1/2

New paper on label-efficient supervised finetuning of LLMs.

We address the expensive prompt annotation cost by humans/proprietary LLMs, saving as much as 50% on FLAN V2.

Paper: arxiv.org/abs/2401.06692
Work led by: Jifan Zhang Yifang Chen Gantavya Bhatt Arnav Das
1/

Why Decision Transformer? It doesn't require the Bellman Completeness -- a strong assumption needed by Q-learning

I'm recruiting PhD students Allen School UW NLP (bdata.uw.edu). Focus areas include Human-AI collaboration, language agents, LLM safety & applications to mental health, social sciences, education.

Apply here: cs.washington.edu/academics/phd/…

UW Data Science
@UW_iSchool
HCI & Design at UW

Zihan will present our work on optimal sample complexity for reinforcement learning: arxiv.org/abs/2307.13586

We proved Q* is hard 4 years ago 🤷‍♂️

#UWAllen is hiring! Join our outstanding scholarly community at University of Washington shaping the future of computing—and having fun while doing it! Priority will be given to faculty applications received by Nov. 13 (teaching track) & Nov. 15 (tenure track). Please share! cs.washington.edu/faculty_candid…

Now Scan&Snap has a follow-up!

1/ We introduce JoMA (arxiv.org/abs/2310.00535), a joint dynamics for MLP lower and self-Attention layers, in order to better understand (1) how multilayer Transformer with MLP nonlinearity works, and (2) qualitatively explain how hierarchical

Want to get model-based RL to work in diverse, dynamic scenes? Check out Chuning Zhu's latest work (RePo) on model-based reinforcement learning without reconstruction, where we show how to learn world models that scale to dynamic, multi-task environments. A 🧵(1/6)

Super exciting work by Allen School's Sheng Wang's lab published in Nature Machine Intelligence Nature Machine Intelligence! 👇
nature.com/articles/s4225…

Introducing our framework and benchmarks for label-efficient learning.

Evaluations of large pretrained models, Semi-SL and active learning have mostly stayed isolated. LabelBench combines all these mutually beneficial techniques to examine the best possible label-efficiency 1/

Our attempt to open the black box of 1-layer transformer training dynamics.