CFP: The 'AI4Research' workshop co-loc. w/ IJCAI'24 is calling for papers about leveraging AI for the research of any discipline. Link: ai4research.github.io Deadline: May 20. Coorg. w/ Rui Zhang Lifu Huang Najoung Kim 🫠 Wei Xu 🌴Muhao Chen🌴, Heike Adel and Jihyun (Janice) Ahn

I had a great time visiting UCLA and USC during the spring break to give talks on 'Amazing Multilingual Capabilities and Concerning Cultural Biases in Large Language Models'.

Thanks to Jesse Thomason Violet Peng Kai-Wei Chang Robin Jia for hosting!

Lots of people in CS are (almost surely) GPT-ing their peer reviews
arxiv.org/abs/2403.07183

One interesting finding of our paper is that a potential cause of cultural biases in LLMs is the heavy use and upsampling of Wikipedia data in pre-training of almost all LLMs.

More Western cultural concepts are written about in non-Western languages, than the other way around.

Thank Michael Nuñez for interviewing us and covering our research in this wonderful VentureBeat article 'LLMs exhibit significant Western cultural bias, study finds', in which I was able to discuss my thoughts on LLM localization and pre-training --

venturebeat.com/ai/large-langu…

Do LLMs understand new emerging words?

We create Neo-Bench: a new dataset of English neologisms. Pre-trained LLMs (GPT-4, LLaMA-2, OLMo, Mistral) fail to translate and answer questions on neologisms.

Preprint: arxiv.org/abs/2402.12261

Wei Xu Alan Ritter Machine Learning at Georgia Tech Georgia Tech School of Interactive Computing

LLMs can write impressive-looking summaries of technical texts in plain language. But are they factual? This is critical in medicine, and the answer is tricky! Introducing ⚕️FactPICO, the first **expert** evaluation of this, with explanations

Paper: arxiv.org/abs/2402.11456 🧵1/

It is disheartening to see good papers get rejected (e.g., reviewers asked for too much), while the opposite also happens often …

OLMo-7b is finally out 🎉, and we are releasing everything; weights, intermediate checkpoints, training code and logs, training data and toolkit, evaluation and adaptation code and data.

Most of it has been released, and the rest is coming soon. OLMo-65b and Adapted OLMo-7b are…

When we (or LLMs) explain technical texts, those explanations can be vague or omit detail. How can we characterize such information loss and help lay readers intuitively recover it?

Excited to share InfoLossQA! 🧵

Paper: arxiv.org/abs/2401.16475
Website: InfoLossQA.ikim.nrw

The 1st South NLP Symposium will take place on April 5, 2024 (F) at Emory University (Atlanta, GA): southnlp.github.io/southnlp2024/

The event date is rescheduled from March 29 to prevent any conflicts with other venues. The new paper deadline is February 7, 2024: southnlp.github.io/southnlp2024/c…

We kicked off this year’s ML seminar at Georgia Tech Machine Learning at Georgia Tech with Scott Yih Scott Yih from Meta AI at Meta.

A very nice read!

I hope someone can do one for #iclr20204 😄… a non-trivial task.

This is a bit belated, but congrats to Fan Bai for defending his Ph.D. and graduating. Here my group (together with Wei Xu's students) drinking champagne and eating oysters to celebrate 🦪🍾