Randall Balestriero
@randall_balestr
AI Researcher: From theory to practice (and back)
Postdoc @MetaAI with @ylecun
PhD @RiceUniversity with @rbaraniuk
Masters @ENS_Ulm @Paris_Sorbonne
246 Following 6.9K Followers
LeJEPA already identified the world's latents--up to some mixing. There is now a very simple post-training recipe to undo that ambiguity and to recover the disentangled world's features! What a time to be Gaussian :)
📄
Show more
Is it possible to provably recover individual latent variables of the true world, even without reconstruction (e.g., JEPA)?
Yes, with DSReg, a simple regularization that can be applied post hoc to your pretrained model!
Show more
Financial data is RIPE for JEPAs... noise, non-stationarity, unknown ground truth states... yet data and evaluation is not readily available! We resolve that in our latest NeurIPS paper `Towards Financial World Modeling' TLDR: invest with LeJEPA ;)
📄 🧵⬇️
Show more
Self-supervised learning is the engine behind modern computer vision.
In practice, this used to translate to "let's wait around for the next DINO".
But not anymore: now it's LeJEPA's turn, and you can finally pretrain SSL models on your own data.
Here's an overview ⬇️ 1/n
Show more
Introducing H-JEPA: the first end-to-end learned hierarchical world model for long-horizon visual planning!
Built atop SIGReg/LeWM, H-JEPA provides a stable training recipe where semantic abstractions naturally emerge across levels...
Paper+Code: 🧵👇
Show more
LeWM applied SIGReg per time-step, so nothing prevents temporal collapse. If your data has such slow features, simply apply SIGReg on the (N*T, D) tensor. We showed this solves your problem here:
And @vlad_is_ai already spoked about JEPA+Slow features 4 years ago here
Show more
JEPA has a huge blind spot no one talks about: it’s starving for stable, slow features.
Add a SINGLE irrelevant patch, keep it fixed within each episode, and planning falls apart. Make it unpredictable, and performance almost returns to normal. Wild!
More in a blog post soon.
Show more
Delighted to be speaking about our recent progress on JEPAs and on the need for more principled solutions that handle data imbalance, noise, multimodality, reasoning, ...! Happening at MICCAI-FOMO26 in 5min! And thank you for having me as keynote today!
Show more
stable-worldmodel got into neurips 2026!
it's basically everything we keep rebuilding for world model research (data loading, baselines, planning, evals) in one place so nobody has to do it again
Show more
JEPA-style world models are biased towards learning slow, simple features.
We're releasing MotionJEPA, built for learning a balanced representation of both static and dynamic features.
Show more
Close to what we are setting up with two-phase review with quick rejection phase 1 (no rebuttal). And phase 1 reviewers need to validate/rebut an AI generated review. That, and all submissions are public, non-official reviewers can also comment live
Show more
Let’s say 50,000 ICLR submissions × 3 reviews × 2 hours = 300,000 hours, or 150 person-years of expert labor.
Let’s use AI to fight AI slop. I would suggest adding an AI-only screening round, then send only the papers that pass to human reviewers.
Let’s not break the system.
Show more
Check out ICWM to get some sanity back. Deadline in ~2months. We are still tuning things (moving to single blind)... but you can already see our plan in the cfp page. Bloated generalist conferences with rotating leadership can't keep up....
Show more
Sincere advice to paper authors and conference organizers: please cancel ICLR 2027, and work on a policy that restores the prestige that a top-tier publication once held.
The International Conference on World Modeling (ICWM) is single blind and entirely accessible to anyone (like ICLR at the beginning) for that reason. Authors doing AI slop will be publicly known to do so. Anonymous reviewers stay without peer-pressure
Show more
@AnkaReuel Wait, ICLR is double blind even for senior roles, no? Are you referring to other conferences?
I agree we need to penalize slop, but I'm not sure removing blind reviewing is the best option.
Show more
Can't wait for LeDINO 🦖
The DINO team at FAIR Paris is hiring a PhD student to push the frontier of visual representation learning. We offer hard research topics, a research and engineering team supporting you, and a stimulating PhD community. Help us shape the next generation of DINO models! Details👇
Show more
Our 4th world modeling workshop will be looking at *Physics*! Happening at the Aspen Center for Physics end of February (
If you are interested in presenting, submit your abstract here:
Deadline on October 9th, 1 page Abstract!
Show more
Will be speaking at Harvard today at 9:45 about world models, Le* family and why we need more theory and mathematics to advance JEPAs!
Come say hi if you are around!
If you missed our opening remarks:
- 4th world modeling workshop: Aspen, CO, February, world models for physics,
- 1st world modeling conference: Bay Area, CA, May,
In collaboration with @ylecun @LambdaAPI @amilabs and more TBA!
Show more
We are 18 hours away from the World Modeling Workshop!
Quick things:
- for the latest infos and the YouTube livestream link (free and open to everyone)
- >= 2 annoucements during opening remarks
- amazing lineup of speakers/attendees
See you soon!
Show more
SIGReg for pretraining Video Foundation Models! Our LeVJEPA opens many doors...
- stable recipe with a simple loss (sigreg + prediction)
- no tubelet, frame aggregation, EMA, stop-gradient, ....
- 20X more FLOP efficient than VJEPA1/2 pretraining
- open source + reproducible
Show more
It's a bird, it's a plane, it's a JEPA! Congrats on that great work that brought SIGReg and JEPAs to the sky--in the real world! Check out the paper!
