Happy Halloween!

There’s plenty of evidence for political bias in LLMs, but very few evals reflect realistic LLM use cases — which is where bias actually matters. IssueBench, our attempt to fix this, is accepted at TACL, and I will be at #EMNLP2025 next week to talk about it! New results 🧵

I will be giving a talk at @eth-ai-center.bsky.social next week, on RLVR for verifiable instruction following, generalization, and reasoning! 📢 Join if you are in Zurich and interested in hearing about IFBench and our latest Olmo and Tülu works at @ai2.bsky.social

"Although I hate leafy vegetables, I prefer daxes to blickets." Can you tell if daxes are leafy vegetables? LM's can't seem to! 📷 We investigate if LMs capture these inferences from connectives when they cannot rely on world knowledge. New paper w/ Daniel, Will, @jessyjli.bsky.social

💡We kicked off the SoLaR workshop at #COLM2025 with a great opinion talk by @michelleding.bsky.social & Jo Gasior Kavishe (joint work with @victorojewale.bsky.social and @geomblog.bsky.social ) on "Testing LLMs in a sandbox isn't responsible. Focusing on community use and needs is."

Hi #COLM2025! 🇨🇦 I will be presenting a talk on the importance of community-driven LLM evaluations based on an opinion abstract I wrote with Jo Kavishe, @victorojewale.bsky.social and @geomblog.bsky.social tomorrow at 9:30am in 524b for solar-colm.github.io Hope to see you there!

Now accepted to #neurips25 datasets & benchmarks! See you in San Diego! 🥳

🚀 Can open science beat closed AI? Tülu 3 makes a powerful case. In our new #WiAIRpodcast, we speak with Valentina Pyatkin (@valentinapy.bsky.social) of @ai2.bsky.social and the University of Washington about a fully open post-training recipe—models, data, code, evals, and infra. #WomenInAI 1/8🧵

"𝐋𝐋𝐌 𝐏𝐨𝐬𝐭-𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠: 𝐎𝐩𝐞𝐧 𝐒𝐜𝐢𝐞𝐧𝐜𝐞 𝐓𝐡𝐚𝐭 𝐏𝐨𝐰𝐞𝐫𝐬 𝐏𝐫𝐨𝐠𝐫𝐞𝐬𝐬 " 🎙️ On Sept 17, the #WiAIRpodcast speaks with @valentinapy.bsky.social (@ai2.bsky.social & University of Washington) about open science, post-training, mentorship, and visibility #WiAIR #NLProc

With fresh support of $75M from NSF and $77M from NVIDIA, we’re set to scale our open model ecosystem, bolster the infrastructure behind it, and fast‑track reproducible AI research to unlock the next wave of scientific discovery. 💡

On my way to Oxford: Looking forward to speaking at OxML 2025

🔈For the SoLaR workshop @COLM_conf we are soliciting opinion abstracts to encourage new perspectives and opinions on responsible language modeling, 1-2 of which will be selected to be presented at the workshop. Please use the google form below to submit your opinion abstract ⬇️

I had a lot of fun contemplating about memorization questions at the @l2m2workshop.bsky.social panel yesterday together with Niloofar Mireshghallah and Reza Shokri, moderated by @pietrolesci.bsky.social who did a fantastic job! #ACL2025

I'll be at #ACL2025🇦🇹!! Would love to chat about all things pragmatics 🧠, redefining "helpfulness"🤔 and enabling better cross-cultural capabilities 🗺️ 🫶 Presenting our work on culturally offensive nonverbal gestures 👇 🕛Wed @ Poster Session 4 📍Hall 4/5, 11:00-12:30

🔥tokenization panel!

why is vancouver sushi so good? 🤤 (vancouver food in general actually)

This week is #ICML in Vancouver, and a number of our researchers are participating. Here's the full list of Ai2's conference engagements—we look forward to connecting with fellow attendees. 👋

I'll be at ICML in Vancouver next week! #ICML2025 You can find me at the following: - giving an invited talk at the "Models of Human Feedback for AI Alignment" workshop - giving an invited talk at the "AI for Math" workshop I'll also present these two papers ⤵️

In Geneva🇨🇭to attend the International Open-Source LLM Builders Summit and present OLMo and Tülu!

💡Beyond math/code, instruction following with verifiable constraints is suitable to be learned with RLVR. But the set of constraints and verifier functions is limited and most models overfit on IFEval. We introduce IFBench to measure model generalization to unseen constraints.

plus, some fun RL experiments

This new benchmark created by @valentinapy.bsky.social should be the new default replacing IFEval. Some of the best frontier models get <50% and it comes with separate training prompts so people don’t effectively train on test. Wild gap from o3 > Gemini 2.5 pro of like 30 points.

Introducing IFBench, a benchmark to measure how well AI models follow new, challenging, and diverse verifiable instructions. Top models like Gemini 2.5 Pro or Claude 4 Sonnet are only able to score up to 50%, presenting an open frontier for post-training. 🧵

Check out our take on Chain-of-Thought. I really like this paper as a survey on the current literature on what CoT is, but more importantly on what it's not. It also serves as a cautionary tale to the (apparently quite common) misuse of CoT as an interpretable method.

🚨Submission deadline extended to June 27th AoE!🚨 Our reviewer interest form is also open! See below for more details👇

Interested in shaping the progress of responsible AI and meeting leading researchers in the field? SoLaR@COLM 2025 is looking for paper submissions and reviewers! 🤖 ML track: algorithms, math, computation 📚 Socio-technical track: policy, ethics, human participant research

I collected some folk knowledge for RL and stuck them in my lecture slides a couple weeks back: web.mit.edu/6.7920/www/l... See Appendix B... sorry, I know, appendix of a lecture slide deck is not the best for discovery. Suggestions very welcome.

i created a gist with some non default llm courses gist.github.com/yoavg/95bbc5...

I’m thrilled to share RewardBench 2 📊— We created a new multi-domain reward model evaluation that is substantially harder than RewardBench, we trained and released 70 reward models, and we gained insights about reward modeling benchmarks and downstream performance!

RewardBench 2 is here! We took a long time to learn from our first reward model evaluation tool to make one that is substantially harder and more correlated with both downstream RLHF and inference-time scaling.

Happy to announce the first workshop on Pragmatic Reasoning in Language Models — PragLM @ COLM 2025! 🎉 How do LLMs engage in pragmatic reasoning, and what core pragmatic capacities remain beyond their reach? 🌐 sites.google.com/berkeley.edu/praglm/ 📅 Submit by June 23rd

Our discussion period just started. Authors, please read our instructions carefully. We require responses by June 2. But, what you really want to hear about is stats .... right? -> 🧵

Is very classic that most people don't know the Tulu 3 paper coined the term RLVR

📢 The SoLaR workshop will be collocated with COLM! @colmweb.org SoLaR is a collaborative forum for researchers working on responsible development, deployment and use of language models. We welcome both technical and sociotechnical submissions, deadline July 5th!

How do language models generalize from information they learn in-context vs. via finetuning? In arxiv.org/abs/2505.00661 we show that in-context learning can generalize more flexibly, illustrating key differences in the inductive biases of these modes of learning — and ways to improve finetuning. 1/

Accepted to #ICML2025! 🥳 See you in Vancouver this summer!

Accepted to #ICML2025 🤩

Very interesting oral history -- interviews with some top NLP folks on the effects of GenAI on their field: www.quantamagazine.org/when-chatgpt...

✈️ I'm in New Mexico for #NAACL2025! Would love to meet and chat about training data, factuality, transparency, robustness, doing a PhD in AI🤖, or anything else. I love to make friends, please say hi if you see me! 🌵🌞 And check out our work at NAACL👇

Heading to NAACL? With "verification being the key to AI" you should go to the poster session Friday, 9-10:30am to chat with my star colleagues @valentinapy.bsky.social + @jacobcares.bsky.social about RewardBench (and really RewardBench 2, evaluation, and reward models in post-training).

I asked "on the other platform" what were the most important improvements to the original 2017 transformer. That was quite popular and here is a synthesis of the responses:

Valentina and I will be presenting RewardBench at NAACL! Come say hi at the poster session on Friday and we can chat about reward models, staying up for 30 hours straight to rapidly reset from Singapore time, and more 🏜️

I'll be at #NAACL2025: 🖇️To present my paper "Superlatives in Context", showing how the interpretation of superlatives is very context dependent and often implicit, and how LLMs handle such semantic underspecification 🖇️And we will present RewardBench on Friday Reach out if you want to chat!

🟢 Announcing the #NAACL2025 Award Winners! The Best Paper and Best Theme Paper winners will present at our closing session 2025.naacl.org/blog/best-pa...

I'm in Singapore for ICLR to present this paper: Tomorrow, April 26th, 10-12:30 in Hall 3+2B #236 Come check it out! arxiv.org/abs/2504.12459

💡 New ICLR paper! 💡 "On Linear Representations and Pretraining Data Frequency in Language Models": We provide an explanation for when & why linear representations form in large (or small) language models. Led by @jackmerullo.bsky.social, w/ @nlpnoah.bsky.social & @sarah-nlp.bsky.social

I'm on my way to ICLR! Let me know if you want to meet and/or hang out 🥳

Jack Merullo, Noah A. Smith, Sarah Wiegreffe, Yanai Elazar: On Linear Representations and Pretraining Data Frequency in Language Models https://arxiv.org/abs/2504.12459 https://arxiv.org/pdf/2504.12459 https://arxiv.org/html/2504.12459

WildBench got accepted to ICLR and will be presented next week in Singapore! 🦁

Everything Marius writes is worth reading! 👏