SmolLM3 uses the APO preference loss! Karel D’Oosterlinck great to see APO getting more adoption!

this paper will be presented at COLM later this year! looking back, i'm glad i tried something slightly out of my normal range in interp. ultimately, i feel that real-world models are much messier than can be satisfactorily explained via behaviour -- we must open the blackbox

i forgot the whole point of saying you're at a conference is to advertise your poster please come check out AxBench by Zhengxuan Wu* me* et al. on Tuesday, 15 July at 11 AM - 1:30 PM

Overdue job update -- I am now: - A Visiting Scientist at Schmidt Sciences, supporting AI safety and interpretability - A Visiting Researcher at the Stanford NLP Group, working with Christopher Potts I am so grateful I get to keep working in this fascinating and essential area, and

ICML ✈️ this week. open to chat and learn mech interp from you. Aryaman Arora and i have cool ideas about steering, just come to our AxBench poster. new steering blog: zen-wu.social/steer/index.ht… 中文: zen-wu.social/steer/cn_index…

Excited to take part in TEDAI San Francisco this fall!

The #SIGIR2025 Best Paper just awarded to the WARP engine for fast late interaction! Congrats to Luca Scheerer🎉 WARP was his ETH Zurich MS thesis, completed while visiting us at @StanfordNLP. Incidentally, it's the fifth Paper Award for a ColBERT paper since 2020!* Luca did an

Introducing ColQwen-Omni, a 3B omnimodal retriever that extends the ColPali concept of multimodal retrieval with late interaction to audio chunks and short videos, with no performance degradation on visual document retrieval wrt our best models! (1/N)

We're presenting the Mechanistic Interpretability Benchmark (MIB) now! Come and chat - East 1205. Project led by Aaron Mueller Atticus Geiger Sarah Wiegreffe

imagine the lucy benchmark -- can u get lucy to do IMO problems? can u get lucy to draw a well-proportioned person? can u get lucy to stop scrolling labubus on tiktok? truly difficult, next-level stuff

How does prompt optimization compare to RL algos like GRPO? GRPO needs 1000s of rollouts, but humans can learn from a few trials—by reflecting on what worked & what didn't. Meet GEPA: a reflective prompt optimizer that can outperform GRPO by up to 20% with 35x fewer rollouts!🧵

Paper: arxiv.org/abs/2507.19457 GEPA will be open-sourced soon as a new DSPy optimizer. Stay tuned! Incredibly grateful to the wonderful team Shangyin Tan Dilara Soylu Noah Ziems Rishi Khare Krista Opsahl-Ong Arnav Singhvi Herumb Shandilya Michael Ryan @ ACL 2025 🇦🇹 Meng Jiang Christopher Potts

It's so funny that prompt optimization is turning out to be more important than fine tuning. Given every ML engineer told me four years ago prompting was irrelevant.

Mike Taylor The better we make models at following prompts, the more impact prompts should have, right?

📅 10-Year ToT Award (2015) Thang Luong, Hieu Pham & Christopher D. Manning: “Effective Approaches to Attention-based Neural Machine Translation” EMNLP 2015 🔗 aclanthology.org/D15-1166/ A milestone in neural MT and attention mechanisms. 🔁🧠

An alternative title for this classic paper could be "Attention seems like it might be all you need?"

The Stanford NLP Group founders won both 2025 ACL 2025 test of time awards: ▪25 yrs: Gildea & Dan Jurafsky, Automatic Labeling of Semantic Roles aclanthology.org/P00-1065/ ▪10 yrs: Thang Luong, Hieu Pham & Christopher Manning, Effective Approaches to Attention-based NMT aclanthology.org/D15-1166/