if you value intelligence above all other human qualities, you’re gonna have a bad time

🚀 How far can RL scaling take LLMs? Drop ProRLv2! 🔥With ProRLv2, we keep expanding LLM’s reasoning boundaries through 3,000+ RL steps over 5 domains and set a new state-of-the-art 🌟 among 1.5B reasoning models. 🔗 Full blog: research.nvidia.com/labs/lpr/prorl… 🤗Open model:

Actually, we implemented this kind of capability in our AdaCtrl paper three months ago by injecting difficult-aware tags (i.e., easy, hard, adaptive) to trigger different reasoning behaviors of LLMs. Paper: arxiv.org/pdf/2505.18822

Google DeepMind just dropped a Nature paper: “A personal health large language model for sleep and fitness coaching.” > Gemini is better that doctors and trainers at sleep and fitness > paper shows huge benefit of AI personalization for health and long-form coaching Honestly

Thanks so much again to the IJCAI25 organizers for the opportunity to share our work on AI+Science! I’m grateful to work with our amazing collaborators and students at Molecule Maker Lab Institute

📣 Our paper is accepted to Findings of EMNLP 2025! Many thanks to all the co-authors! 🌍 Math modeling is the perfect lens for agents to approach the real world challenges. Come and check how we do it: arxiv.org/pdf/2505.15068

Congratulations to everyone! This was the *only paper* I worked on overnight during my PhD, but fortunately, I had a group of friends by my side. It is truly a remember memory.

Thanks again to the organizers! What inspired me most was Yoshua Bengio’s incredible kindness and humility. He welcomed different opinions with genuine openness. I could also see the same spirit of kindness and deep care for humanity reflected in his mentees, like Kyunghyun Cho

For agents to improve over time, they can’t afford to forget what they’ve already mastered. We found that supervised fine-tuning forgets more than RL when training on a new task! Want to find out why? 👇

I'm hiring 1-2 new postdocs to work on AI for Scientific Discovery (especially on drug discovery and material discovery) and Science-Inspired AI (especially on scientific foundation models). Please drop me an email if you are interested or know someone who might be a great fit!

When the agent can win Nobel Prize in Physics? 🧐

Excited to shared that ToolRL is accepted to NeurIPS Conference🎉 I was watching Denny Zhou’s Simon Institute talk live, where in one slide, he defined RL Fine-Tuning as “directly optimize what you want”. This very simple reframe completely shifted my perspective on training and the

📢 Early submission for ACL 2026 via ARR Oct cycle is available to support authors facing potential visa delays. 📝 Early invitation letters possible for submit-ready work (not acceptance guarantees). ⚠️ Preliminary work risks rejection & Jan-cycle ineligibility. #NLProc #ARR

> working on semantic parsing in PhD > didn't even have its own track at ACL > it's a dead area, people say > had ~100 citations when graduating > but natural language programming is always the dream > 'Let machines understand human thinking. Don’t let humans think like machines'

I'm seeing quite a bit of comment about this, so I want to make a couple of points. I'm not owed eternal agreement from any actor who once played a character I created. The idea is as ludicrous as me checking with the boss I had when I was twenty-one for what opinions I should

Such a good place to spend the weekend with the family.

Safety is the most important part if we want to use agent in daily life.

Multimodal conversational agents struggle to follow complex policies, which also impose a fixed computational cost. We ask: 👉 How can we achieve stronger policy-following behavior without having to include policies in-context? 🌐: mikewangwzhl.github.io/TriMPI/ 🧵1/3

Super proud of this work Zhenhailong Wang did at Amazon

🚀 Introducing LabOS: The AI-XR Co-Scientist A system that sees, understands, and works with humans in real-world labs. 👁️ Egocentric vision & extended reality 🧠 LLM reasoning & hypothesis generation 🤖 Real-time guidance & multi-modal human-AI collaboration From observation →