Today we're launching PRISM, a new resource to diversify the voices contributing to alignment. We asked 1500 people around the world for their stated preferences over LLM behaviours, then we observed their contextual preferences in 8000 convos with 21 LLMs arxiv.org/abs/2404.16019

So much FOMO missing #ICLR2024 this year, but if you're curious what our research team has been up to, try to catch Candace Ross and/or Melissa Hall !

🚀Thrilled to launch the Workshop on the Next Generation of AI Safety at #ICML2024! Dive into the future of AI safety. CFP & more details 👉 icml-nextgenaisafety.github.io #NextGenAISafety #ICML2024

New week, new blog! In collab with MLCommons, we wrote up a summary of the PRISM alignment project, covering what motivated us to collect feedback from 1,500 humans around the world, how we collected it & which challenges we encountered on the road 🙇‍♀️ mlcommons.org/2024/05/prism/

Very excited to share this amazing resource, congrats to Florian and co.! 🎉

🚨 New from Meta AI 🚨 Want to learn more about VLMs? Ask Florian Bordes from AI at Meta about their latest paper “An Introduction to Vision-Language Models”. This week’s paper of the week 👇 alphaxiv.org/abs/2405.17247

🌎Introducing LINGOLY, our new reasoning benchmark that stumps even top LLMs (best models only reach ~35% accuracy)🥴 In a colab between University of Oxford, Stanford University and UK Linguistic Olympiad puzzle authors, we stress test LLMs on over 90 low-resource and extinct languages...

Check out our new work featured below on measuring geographic disparities in image generations, led by Abhishek Sureddy, Dishant, & Nandhinee Periyakaruppa, along with coauthors Oindrila Saha, Adina Williams, Adriana Romero-Soriano, Polina Kirichenko, Melissa Hall! 🎉

I've got a cool opportunity to share! We're looking for folks to create safety prompts for our upcoming safety benchmark MLCommons AI Safety Initiative, and we have some funding to share with qualified orgs! Check it out or sign up here by Jul 19: mlcommons.org/ai-safety/ai-s…

🚨There is serious lack of robustness with MMLU! In our new work we find that “Changing Answer Order Can Decrease MMLU Accuracy” and the accuracy of top models can drop by 10-20%📉 This means leaderboards might not be as reliable as we thought! 📄arxiv.org/abs/2406.19470 (1/N)

🚨Excited to share that the V2 of the social impact paper, led by the incredible Irene Solaiman and Zeerak Talat (زیرک تلت) Zeerak@ mastodon|bsky, is finally out! In this work, we present a guide for evaluating the social impact of Gen AI systems across categories & modalities. 🧵(1/7) arxiv.org/pdf/2306.05949

Video recordings from the #NSF-sponsored workshop New Horizons in Language Science: Large Language Models, Language Structure, and the Cognitive & Neural Basis of language are now publicly available! All videos are linked to from the workshop website: newhorizonsinlanguagescience.github.io 1/

🐘🙋 New Preprint🙋 🐘 Bigger is not always better: The importance of human-scale language modeling for psycholinguistics osf.io/preprints/psya…

I'm at #CogSci2024 giving a poster TODAY from 13-14:15 about the past and future of BabyLM! Come chat about computational models of language acquisition and data-efficient pretraining!

fyi: Meta's funding a few new grants for faculty to create evaluation benchmarks (deadline Sept 7). Could be of interest! #NLProc llama.meta.com/llm-evaluation…

We won OUTSTANDING PAPER at #ACL2024 for our work on evaluating values and opinions in LLMs 🥳 Thank you to the reviewers and awards committee, and again to my amazing co-authors, especially joint first author Valentin Hofmann Check out the poster + paper + full author list below 👇

"Are Female Carpenters like Blue Bananas? A Corpus Investigation of Occupation Gender Typicality" finds relationship between gender mentioning in Wikipedia and Reddit and actual (from BLS-Labor Statistics) femaleness of occupations. (Ju et al, 2023) arxiv.org/pdf/2408.02948 Adina Williams