🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Very grateful to have received an outstanding paper award for our “Do Llamas Work in English?” paper. This paper would not have been possible without 4️⃣ co-firstauthors. Great job everyone! The energetic young upcoming PhD students Venia and Giovanni and the legendary Prof. West!

Exciting 4-yr PhD position pushing the boundaries of LLMs for engineering at the Helmholtz-Zentrum Hereon / TU Hamburg, supervised by the amazing Roland Aydin & co-supervised by me. Applications due on 1 Sep! tinyurl.com/22aawphh #PhDPosition #ML #Engineering #LLM #Hamburg

shout out to my amazing collaborators: Viacheslav Surkov, Veniamin Veselovsky, Giuseppe (Peppe) Russo, Bob West, and Caglar Gulcehre paper link: arxiv.org/abs/2407.06946

Our new mechanistic interpretability work "Activation Scaling for Steering and Interpreting Language Models" was accepted into Findings of EMNLP 2024! 🔴🔵 📄arxiv.org/pdf/2410.04962 Kevin Du, Vésteinn Snæbjarnarson, Bob West, Ryan Cotterell and Aaron Schein thread 👇

Our paper "The AI Review Lottery: Widespread AI-Assisted Peer Reviews Boost Paper Scores and Acceptance Rates" got covered by the The Chronicle of Higher Education. Exciting times for peer-reviewing ICLR 2026 chronicle.com/article/ai-sci…

Disappointed to not be at #EMNLP owing to a dislocated shoulder 😢 Debjit Paul will present our poster on Multilingual Entity Insertion (cf. arxiv.org/pdf/2410.04254). Swing by our poster in session #6 on Wed 13@10:30 EST 🚀 PS: I am hiring PhD students Computer Science at Aarhus University #LLM #GNNs #CSS

Can we understand and control how language models balance context and prior knowledge? Our latest paper shows it’s all about a 1D knob! 🎛️ arxiv.org/abs/2411.07404 Co-led with Kevin Du, as well as Niklas Stoehr, Giovanni Monea, Chris Wendler, Bob West & Ryan Cotterell.

I love this paper, one of my recent favorites — not just because of the amazing animated llama in Julian Minder’s tweet :)

New work with Kristina Gligorić Eric Horvitz Emre Kıcıman Arnaud Chiolero and Ryen White on food choice mimicry at EPFL

"The AI Alignment Paradox," by Bob West (EPFL) and Roland Aydin, says that aligning #AIModels with our values makes it easier for adversaries to misalign them. bit.ly/42E8CmY

Excited to announce the #WikiNLP workshop at ACL 2025! We welcome #NLP contributions to Wikimedia, especially on datasets, and ideas to advance its mission. w/ Isaac Johnson (Wikimedia Foundation), Akhil Arora @ICML’25 🇨🇦, Lucie-Aimée Kaffee, Tiziano Piccardi, Indira Sen See: w.wiki/CumQ

New CACM opinion piece written together with Roland Aydin about what we call the “AI Alignment Paradox”: virtuous AI may be more easily made vicious. Possibly this might be a reason for an intriguing recent result by Owain Evans et al.: x.com/OwainEvans_UK/…

I am recruiting 2 PhD students for Fall'25 Computer Science at Aarhus University to work on bleeding-edge topics in #NLProc #LLMs #AIAgents (e.g. LLM reasoning, knowledge-seeking agents, and more). Details: cs.au.dk/~clan/openings Deadline: May 1, 2025 Please boost! cc: WikiResearch Pioneer Centre for AI SODAS, Copenhagen (Bluesky: @cphsodas.bsky.social)

New paper: Language models have “universal” concept representation – but can they capture cultural nuance? 🌏 If someone from Japan asks an LLM what color a pumpkin is, will it correctly say green (as they are in Japan)? Or does cultural nuance require more than just language?

How do diffusion models create images and can we control that process? We are excited to release a update to our SDXL Turbo sparse autoencoder paper. New title: One Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models Spoiler: We have FLUX SAEs now :)

This work got accepted to ACL 2025 main! 🎉 In this updated version, we extended our results to several models and showed they can actually generate good definitions of mean concept representations across languages.🧵

Very exciting work out of my lab at EPFL Computer and Communication Sciences, led by Saibo-Creator — zip2zip, a method for allowing LLMs to work directly in compressed token space.