Furong Huang

Last month we booted up our robotics lab (see the boot‑up story here 👉x.com/furongh/status…). Today: TraceGen is out — our first product. We’ve been chasing a simple idea with big consequences: the small‑data problem in robotics is really a variation problem. Different bodies, cameras, and scenes fragment experience. So rather than learn the look of the world, learn the shared, scene‑centric 3D structure of motion—the where + how that transfers across embodiments. This is why we train on web‑scale video across embodiments (human↔robot, robot↔robot; new cameras, new scenes) to build a transferable motion prior. Focus on geometry that matters for manipulation; treat appearance as incidental. Variation stops being a tax and becomes fuel. Alongside the release, we’re opening traceforge—our dataset and tooling for working in this “trace” view—so others can reuse in‑the‑wild video without wrestling pixels or prose. If our vision for physical intelligence resonated with you—structure over surface, reuse over recollect—we’d love feedback, bugs, and collaborations. 🌐 Website: tracegen.github.io 🔧 Data and Tooling: TraceForge and TraceGen 📄 Paper: arxiv.org/abs/2511.21690 See more details here: x.com/JayLEE_0301/st… #EmbodiedAI #RobotLearning #WorldModels #CrossEmbodiment

Huge thanks to the team Anirudh Satheesh, Michael-Andrei Panaitescu-Liess @michael_panaite, Andrew Xu, Georgios Milis, Heng Huang, Zikui Cai @zikuicai and myself @furongh 💐

Project Site: …tional-adversarial-training.github.io/Code Repository: github.com/Asatheesh6561/… Models & Datasets: huggingface.co/collections/as…

Everyone talks about watermarking AI images. But after WAVES Bench + our NeurIPS “Erasing the Invisible” competition, one thing became clear: watermarks don’t fail in the lab. They fail after the real world touches them. I wrote about our new paper, CAT, and why robust watermarking needs to train against adaptive compositional attacks — not just random corruptions. x.com/furongh/status…

Check out my latest article: When Students Shape the Science: Reflections on Data, AI Governance, Alignment, and World-Modeling for Physical AI linkedin.com/pulse/when-stu… via @LinkedIn

Excited to share that our paper: “Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling” has been selected as an Oral at #ACL2026 🎉 #ACL Reward models are increasingly the hidden control system behind modern AI alignment. But what happens when the reward model itself gets hacked? In this work, we train reward models to actively discover their own blind spots through adversarial failure discovery, improving robustness against reward hacking and distribution shifts. This is part of a broader direction my lab has been exploring recently on mitigating reward hacking, robust alignment, and building AI systems that can reason about their own failures rather than merely optimize superficial rewards. Paper: arXiv link arxiv.org/abs/2507.06419 Huge credit to @PankayarajP who made this possible. He is on the job market this season!! Stay tuned — we’ll share a deeper technical blog post soon.

I'm definitely against this specific method. This particular method has no principled way whatsoever of guaranteeing that the English text inside the autoencoder reliably and faithfully corresponds to what the model is "actually thinking." Optimizing their encoder and decoder jointly for reconstruction does not put any optimization pressure on anything related to making sure that the intermediate text between the encoder and decoder has the same meaning to the decoder as its meaning in English. The fact that they had to use a KL divergence penalty calculated using a text model to make sure that the English intermediates were readable is pretty strong evidence that their method is fundamentally ill-equipped to produce faithful explanations. This issue with this autoencoder approach is not something that you could solve by modifying the method. It would be a category error to have optimism about natural language autoencoders being reliably able to produce faithful explanations.

@henrytdowling Thanks Henry, I will!

@furongh hit me up if you'd like to come by an office in soma where a bunch of startups + researchers work!

Summer itinerary update ☀️ 📍June 7–15: Bay Area Looking for kind souls to give me lab/company tours so I can respectfully geek out and admire what everyone is building 👀 📍June 16–July 5: China Would love to visit labs, startups, and research groups there too! 📍July 5–11: ICML in Seoul 🇰🇷 Come say hi if you’re around! Always excited to meet people working on weird, ambitious, or slightly dangerous AI ideas™ In exchange, I can offer: • caffeinated research rambling • unsolicited opinions about reasoning models • and an enthusiastic download of what we’ve been working on lately Current obsessions include: 🧠 How to actually use test-time compute wisely when information is flying everywhere 🤖 Physical intelligence and embodied agents 🌍 Building world models with symbolic structure that don’t collapse… but also don’t become stubborn brute-force memorization machines Basically: How do we make models think more intelligently, instead of just thinking longer? If anyone wants to chat research / robotics / reasoning / agents / world models / alignment / future AI systems, I’d genuinely love to connect. Also yes, I am absolutely the kind of person who visits a lab and immediately asks: “okay but what weird thing are you secretly most excited about?” PS: Since Attention is All You Need, I am posting this picture from Rio, Brazil, (stolen) from my student Weize Liu @WeizeLiu1115 😬

@zhuokaiz A naive prompt "Could you photoshop me into this picture? I am providing the picture and a picture of me." gives me this 👇 LOL

@furongh Guess I wasn’t expecting ai can understand what “photo shopping me” means and generate with the correct style, amazing!

@zhuokaiz I think AI understood "meme" 😊

Here is the prompt I used: "I am uploading three pictures. One picture is a group of people surrounding Yann LeCun. The other two pictures are me. I want to make a meme on Twitter by photo shopping me into this picture and say something like: I didn’t manage to get into the picture with Yann, but well, with AI, I can do anything!" I think it did a great job understanding what a meme entails, and correctly did the clumsy photoshop!

@furongh It looks more like old-school photoshop than ai 🤣

In case you are curious: this is the original

To draw a packed room for a defense the day before the NeurIPS deadline - and to field the highest number of questions in the Q&A I’ve seen in over 50 defenses during my faculty career – speaks volumes about the researcher. John @jwkirchenbauer, I am truly honored to serve on your committee. Between your Best Paper Award for LLM watermarking and the way you take absolute ownership of your research, your unwavering professionalism is a masterclass for us all. Congratulations! 🌟🎓

Honestly bringing your own lighting is such a clever solution to getting robot policies to work in every home

I am excited to give a talk and participate in a panel today Sunday Apr 26 at the AFAA workshop📍211,: Talk:📅 14:00 - 14:45; Panel: 📅 16:15 - 16:55.

Here are some pictures from the oral session. Thanks @ju_yuanchen for the polished slides!