"LLMs can play games" is a fashionable trend. As we demonstrate in our paper, training on a specific set of games yields higher performance on those games alone. However, this doesn't help models to play unseen games, showcasing their limitations in true instruction following! ↓

Robotics models are increasingly bulky and difficult to run directly on robots. With Remi Cadene and the team LeRobot and Hugging Face we’re changing that. Introducing SmolVLA, a sub-500M VLA designed for efficient training and inference. A thread 🧵

🚨 5 DAYS TO GO! The world’s biggest AI Robotics Hackathon is almost here! 2,000+ builders, coders, dreamers are joining June 14–15 One rule: build, learn and have fun together! Find your local hackathon & team up Register now: forms.gle/NP22nZ9knKCB2K…

Super cool opportunity to work with us on implementing novel Embodied #GenAI to enable human-robot collaboration! Reach out if you want to know more!

This is beautiful. Well done Rajat and the rest of the team. Truly inspiring!

🚨Meet MF²: Movie Facts & Fibs: a new benchmark for long-movie understanding! 🤔Do you think your model understands movies? Unlike existing benchmarks, MF² targets memorable events, emotional arcs 💔, and causal chains 🔗 — things humans recall easily, but even top models like

Reasoning is about variable binding. It’s not about information retrieval. If a model cannot do variable binding, it is not good at grounded reasoning, and there’s evidence accruing that large scale can make LLMs worse at in-context grounded reasoning. 🧵

Current AI “alignment” is just a mask Our findings in The Wall Street Journal explore the limitations of today’s alignment techniques and what’s needed to get AI right 🧵

Any updates on the status of Open Review? I really need to submit my NeurIPS Conference reviews :(

I hope somebody mentioned pixel-based models: arxiv.org/abs/2401.03321 antonio vergari ⚔️ not at #ICML2025

New Anthropic Research: “Inverse Scaling in Test-Time Compute” We found cases where longer reasoning leads to lower accuracy. Our findings suggest that naïve scaling of test-time compute may inadvertently reinforce problematic reasoning patterns. 🧵

The amazing folks at EdinburghNLP will be presenting a few papers at ACL 2025 (ACL 2025); if you're in Vienna, touch base with them! Here are the papers in the main track 🧵

Looking forward to kicking off the day2 at #ACL2025NLP with my keynote! We'll be tackling new frontiers of AI alignment. 🗓️ Tuesday, 9:00 AM 🗣️ "Who's Gold? Re-imagining Alignment for Truly Beneficial AI" Here's a sneak peek of the talk. #AI #AIAlignment #NLProc #ACL2025NLP

At #ACL2025NLP and on the job market (NLP + AI Safety) 💼 It's great to see growing interest in safety/alignment, but we often miss the social context. Come to our Workshop on Online Abuse and Harms Friday to dive deeper into safe safety research! A quiet token from the biggest ACL 2025 ⬇️

Wish I could be doing this. Check this proposal from Mario!

Special thanks to Google DeepMind for inviting me to try out Genie 3. I'm excited to share my thoughts on this early research prototype and also some of my live recordings below: I spent the whole day playing with the system and when it works, it is truly mind blowing🤯. It is

This looks amazing. I wonder how robust it is but seems definitely the way to go for setting up a truly open-ended learning regime where the learning agent is constantly challenged with novel and learnable experiences. Congrats Genie team!

When Sam Altman told me at the AI summit in Paris that they were serious about releasing open-source models & asked what would be useful, I couldn’t believe it. But six months of collaboration later, here it is: Welcome to OSS-GPT on Hugging Face! It comes in two sizes, for both