Our paper CTRL-D is accepted to ACL Findings and will be presented at ACL 2025! 🗓️Poster session: 18:00–19:30 (Level 0 Exhibit Halls X4/X5) I’m sad I can’t be there, but Jordan (@boydgraber) will! You’ll enjoy learning about CTRL-D from him. Now… what is CTRL-D? 🔍

🚨 My last paper as a postdoc is out! Practical Alignment: Why Learning from Human Feedback Alone Is Not Enough for Safe and Helpful AI. This paper started from a simple observation: Humans have incorrect beliefs about the world — and today’s alignment methods pretend they

The toughest moment in a PhD: >spend a year building smth you’re proud of >travel across world with advisors’ support to share it >when your moment finally comes >your mic gets cut because the previous schedule ran late Heartbroken, but thanks to whom stayed for me @NeurIPSConf

Paper primarily from @Princeton and @UofIllinois! 😛

The MTI-LLM Workshop at @NeurIPSConf 2025 is coming next Saturday. We know agents need to work over long horizons. But we still lack a clear roadmap for agents that can work robustly over extremely long horizons. How do we scale RL training algorithms for long-term planning? How

🎉 Stoked this is finally out! Large scale meta learning allowed us to discover an algorithm that outperforms human-crafted ones on a range of RL benchmarks, including ones that we didn’t train on. Check it out 🤗

Attention: Google has dropped a new version of Attention Is All You Need!

Brutal honesty is what you want from a great research advisor

I'm recruiting my first group of PhD students at TTIC! If you're interested, please apply! If you know people who might be interested, please spread the word! Application deadline is Dec 9, 2025, and there is no application fee:

I'm really excited about our new paper!! 📣 'Reinforcement Learning Improves Traversal of Hierarchical Knowledge in LLMs' Contrary to belief that RL ft degrades memorized knowledge, RL-enhanced models consistently outperform base/SFT on knowledge recall by 24pp! RL teaches

On my way to #EMNLP2025 🇨🇳 I’ll be presenting our work (Oral) on Nov 5, Special Theme session, Room A106-107 at 14:30. Let’s talk brains 🧠, machines 🤖, and everything in between :D Looking forward to all the amazing discussions!

cool idea from Meta What if we augment CoT + RL’s token space thinking into a “latent space”? This research proposes “The Free Transformer”, with a way to let LLMs make global decisions within a latent space (via VAE encoder) that could later simplify autoregressive sampling

Lots of work on cross-lingual alignment encourages multilingual LLMs to generalize knowledge across languages. But this push for uniformity creates a tension: what happens to knowledge that should remain local? We look into this trade-off of transfer and cultural erasure:🧵

So proud of Ruijie! He joined my lab as a sophomore and has grown with us ever since. I still remember his very first lab presentation, he wrote every equation by hand in real time. I knew right then he was different. He started with RL theory, explored robustness, then moved

Life update: I’ve successfully defended my PhD thesis today and will soon be joining the GEAR Lab as a research scientist for building humanoid robot foundation model. It’s been such a wonderful journey at Maryland, already starting to miss it!

Today's AI agents are optimized to complete tasks in one shot. But real-world tasks are iterative, with evolving goals that need collaboration with users. We introduce collaborative effort scaling to evaluate how well agents work with people—not just complete tasks 🧵

🧠🚀 Excited to introduce Supervised Reinforcement Learning—a framework that leverages expert trajectories to teach small LMs how to reason through hard problems without losing their minds. 🤯 Better than SFT && RLVR. Read more: https://t.co/taEL8Vk4X5 #llms #RL #reasoning

@karpathy observed LLMs are "silently collapsed...only know 3 jokes". We prove this is mathematically inevitable due to RLHF + human psychology. But these capabilities aren't lost, just hidden – and easily restored. This means AI benchmarks are measuring training artifacts.🧵

We scaled up an "alternative" paradigm in RL: *divide and conquer*. Compared to Q-learning (TD learning), divide and conquer can naturally scale to much longer horizons. Blog post: https://t.co/xtXBzya0bI Paper: https://t.co/nqYkLucsWu ↓

🚀 AI can sketch what it thinks - no external tools, just pure latents! VLMs can see but can’t imagine🧠 What if they can sketch their thoughts directly?🎨 Meet Latent Sketchpad — the next leap for MLLMs, giving models an internal visual scratchpad to truly think in images.

New Nvidia paper shows how a single LLM can teach itself to reason better. It creates 3 roles from the same model, a Proposer, a Solver, and a Judge. The Proposer writes hard but solvable questions that stretch the model. The Solver answers those questions with clear steps and