This is exciting! Two independent groups have now confirmed the presence of self-bias in various tasks performed by large language models. Our paper, 'Perils of Self-Feedback' (arxiv.org/abs/2402.11436), revealed that self-refinement processes may actually intensify the self-bias.

Thank you for the interview Matteo Wong The Atlantic , great to hear diverse opinions about the topic Bonaventure F. P. Dossou Ife Adebara Antonis Anastasopoulos @EACL alexandra birch Sara Hooker Holy Lovenia Mehak Dhaliwal

A wonderful piece which draws on researcher perspectives and communities working on this problem.
David Ifeoluwa Adelani 🇳🇬 Bonaventure F. P. Dossou Ife Adebara Antonis Anastasopoulos @EACL alexandra birch Holy Lovenia Mehak Dhaliwal

+ a spotlight on global multi-year efforts: Cohere For AI w Aya + Masakhane

What do you do when someone questions your work while you're giving a talk?

You do a collaboration!

Coming to NAACL 24: a paper investigating when contested translations matter for text-to-image multilinguality assessment w/ Yiran "Lawrence" Luo from ASU

arxiv.org/abs/2403.11092 1/8

🚀 Exciting News! 🚀

We had our first Real AI group meeting last week, and it was a fantastic start to our journey together! It was amazing to see everyone's enthusiasm and passion for AI. At Real AI, we seek to pioneer the next frontiers of artificial intelligence for science.

🎉 Thrilled to share my first first-author paper on adversarial transfer learning got accepted at #CVPR2024 ! Feel free to check it out: arxiv.org/abs/2312.05716

Brief Summary ⏬

{UCSB|AI2|UW|Stanford|MIT|UofT|Vector|Contextual AI} present a survey on🔎Data Selection for LLMs🔍

Training data is a closely guarded secret in industry🤫with this work we narrow the knowledge gap, advocating for open, responsible, collaborative progress
arxiv.org/abs/2402.16827

In case anyone else is digging through their bookmarks looking for these today, here are the rebuttal guides ;)

docs.google.com/document/d/1cd…

3. MT detection and multi-way parallelism are promising data filtering techniques for improving multilingual data quality. This is also an important issue for AI researchers and internet content consumers as machine-generated content is ever-increasing.

2. The highly multi-way parallel translations are likely machine translations from English to lower-resource languages with significantly lower translation and content quality

Thank you Jules Roscoe VICE News for covering our work! Read at arxiv.org/abs/2401.05749

Key takeaways-
1. A large portion of the internet is translated to many languages (multi-way parallel), with low-resource languages having significantly more translations

Heads up for anyone looking to post an ACL-format preprint to arXiv on Jan 15: that's a US public holiday (MLK day), so there will be no postings that day. If you want to be able to link to your preprint before AOE Jan 15, your arXiv submission deadline is actually THIS Friday!

RAG for LLMs

Great to see an overview of all the retrieval augmented generation (RAG) research that has been happening.

Should be a great read for the end of the year.

arxiv.org/abs/2312.10997…

Join for research in responsible and ethical AI, led by the best advisors!

Interested in AI for Science? Our team Real AI @ UCSB is recruiting!

Haewon Jeong , Yao Qin & I are hiring postdocs & PhD students & visiting researchers 😃

👉Apply to UCSB ECE at ece.ucsb.edu/grad/apply before Dec 15!

Yao & I are also #NeurIPS2023 : feel free to reach out!

In particular, we are working on AI Safety (robustness & fairness), AI for Science (quantum and neuroscience), AI for Healthcare (Diabetes & Women Brain Health). Feel free to reach out if you are interested in one of these topics!

The REAL AI team @ UCSB Real AI @ UCSB is hiring! We have openings for PhD students/Postdoc/visiting researchers! Both Nina Nina Miolane and I @YaoQinUCSD will be at NeurIPS this week. Come to talk to us if you have interests in doing REAL AI research 😊

Conventional #recsys wisdom: 'better to go wide than deep'. Our paper: go infinitely-wide, compute the solution in closed-form with a single hyper-parameter, and considerably beat all SoTA. Furthermore, can you get the same performance with just 500 fake users? Yes! A thread 🧵