600 experiments reveal JEPA world model insights
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Author Haddy Alchaer ran 600 experiments on the LeWM JEPA world model and raised its long-horizon success rate from 22% to 85% without retraining.
The results suggest planning and goal scoring matter more than perfect physics for robot world models.
Background
JEPA (Joint Embedding Predictive Architecture) is a world-model approach proposed by Turing Award winner Yann LeCun of NYU and AMI Labs, in which the model predicts in embedding space rather than generating future pixels. LeWM, a JEPA world model released this year by a team including LeCun, uses Push-T as its main benchmark. Haddy Alchaer's study tests how much LeWM's planning depends on accurate physics versus how the task is set up, amid a broader debate over whether world models should predict pixels or latent representations.
How it unfolded
- 2022Yann LeCun proposed JEPA, the Joint Embedding Predictive Architecture.
- 2025Meta released V-JEPA 2, pre-trained on more than one million hours of internet video and post-trained on under 62 hours of robot data.
- Nov 9, 2026The Register and LavX News reported a pre-print, "Counterfactual World Models: Natura Non Facit Saltus," by Randall Balestriero, Kerda Varaku and Yann LeCun.
- Nov 2026Haddy Alchaer published "After 600 experiments, here's what I discovered about JEPA world models," reporting 600 experiments on LeWM.
Key terms
- world model
- An internal representation of external reality that the mind uses to anticipate events and guide reasoning and decisions.
- embedding space
- A mathematical space in which data points are represented as vectors preserving structural relationships.
Source: Wikipedia
Who’s saying what
- Author
- Alchaer says LeWM's predictions are already as good as perfect physics for planning along trained paths, and that changing how the goal is scored raised success by 60 points, three times as much as swapping in perfect physics.
- Expert
- LeCun, per Alchaer's account, argues generative world models that predict in pixels are wasted effort because most pixels don't matter for deciding what to do and many can't be predicted at all.
- Caution
- RobCo's Felix Frank says latent-space JEPA prediction is more efficient but harder to debug than pixel-space video generation.
Still unverified
The counterfactual world models paper is a pre-print, reported as expected on arXiv this week; RoboticsIntl notes that the background-robustness fix for latent world models remains a benchmark-stage result in simulation.