This one weird* trick will fix all** your LLM RLHF issues! * not weird ** as long as your issues are about how to combine multiple objectives, and avoid reward hacking

The ICML Conference 2024 Ethics Chairs, Kristian Lum and Lauren Oakden-Rayner 🏳️‍⚧️, wrote a blog about the ethics review. Helpful for all authors and reviewers at ICML to better understand the process! medium.com/@icml2024pc/et… . . .

The band is getting back together! Tomorrow, I’m joining William Isaac and so many others I admire on Google DeepMind’s Ethics team to work on AI evaluation. Exciting times ahead…

In case you missed it Bioptimus: We are looking for the best talents (ML/biology/large-scale infrastructure) to join our fantastic technical team @ZeldaMariet fllinares Jean-Philippe Vert 👉 Bioptimus.com/careers

Excited to share our #ICLR2024 paper, focused on reducing bias in CLIP models. We study the impact of data balancing and come up with some recommendations for how to apply it effectively. Surprising insights included! Here are 3 main takeaways.

📢📢 I am looking for a student researcher to work with me and my colleagues at Google DeepMind Zürich on vision-language research. It will be a 100% 24 weeks onsite position in Switzerland. Reach out to me ([email protected]) if interested. Bonus: amazing view🏔️👇

Proxy methods: not just for causal effect estimation! #aistats24 Adapt to domain shifts in an unobserved latent, with either -concept variables -multiple training domains arxiv.org/abs/2403.07442 Tsai Stephen Pfohl Olawale Salaudeen Nicole Chiou (she/her) Kusner Alexander D'Amour (@alexdamour.bsky.social) Sanmi Koyejo

In our new working paper "Dubious Debiasing" presented in today's #CHI2024 HEAL workshop, we argue that modern LLMs like ChatGPT cannot be fair in the ways currently conceived in ML/NLP. We need new context-adaptive methods to tackle the evaluation crisis: heal-workshop.github.io/papers/1_dubio…

LLM best-of-n sampling works great in practice---but why? Turns out: it's the best possible policy for maximizing win rate over the base model! Then: we use this to get a truly sweet alignment scheme: easy tweaks, huge gains w Lin Gui Cristina Garbacea arxiv.org/abs/2406.00832

📣Announcing the 2024 NeurIPS Workshop on Statistical Frontiers in LLMs and Foundation Models 📣 Submissions open now, deadline September 15th sites.google.com/berkeley.edu/b… If your work intersects with statistics and black-box models, please submit! This includes: ✅ Bias ✅

New reqs for low to high level researcher positions: jobs.careers.microsoft.com/global/en/job/… , jobs.careers.microsoft.com/global/en/job/…, jobs.careers.microsoft.com/global/en/job/…, jobs.careers.microsoft.com/global/en/job/…, with postdocs from Akshay and Miro Dudik x.com/MiroDudik/stat… . Please apply or pass to those who may :-)

Excited to announce that our paper, “A toolbox surfacing health equity harms and biases in large language models” is now published with Nature Medicine: nature.com/articles/s4159….

If you're a PhD student interested in interning with me or one of my amazing colleagues at Microsoft Research New England (Microsoft Research New England, Microsoft Research) this summer, please apply here jobs.careers.microsoft.com/global/en/job/…

But is the JEC data that Nate is putting into his regression picking up that variation in any meaningful way? I don't buy it. Because it isn't actually making any attempt to measure state-level differences in inflation. It's just picking up preexisting differences in income.

New work! 💪🏻💥🤯 When Can Proxies Improve the Sample Complexity of Preference Learning? Our paper is accepted at ICML Conference 2025. Fantastic joint work with Daniel Augusto @zhengyan_shi Mengyue Yang Pasquale Minervini is hiring postdocs! 🚀 Alexander D'Amour (@alexdamour.bsky.social), Matt Kusner. 1/n

Sadly I can't make it myself to #ICML2025, but the amazing Alexander D'Amour (@alexdamour.bsky.social) is presenting this at West Exhibition Hall B2-B3 W-816, Tuesday 15 Jul 11am - 1.30pm PDT! Come check it out and talk about LLM reward hacking, sample efficiency and low-dimensiomal adaptation with him!

Dusting this example off again in light of some recent education funding discourse