600 Experiments on JEPA World Models: Key Findings
Original: After 600 experiments, here's what I discovered about JEPA world models
Why This Matters
Shows JEPA planning gains come from architecture use, not just model quality.
Researcher Haddy Alchaer ran 600 experiments on LeWM, a JEPA world model, finding four new planning methods that raised long-horizon task success from 22% to 85%—without any retraining.
JEPA (Joint-Embedding Predictive Architecture) is Yann LeCun's alternative to generative world models. Instead of predicting future pixels, JEPA encodes observations into compact embeddings and plans in that compressed space—only modeling what matters for action. This year, LeCun's team released LeWM, a JEPA model benchmarked on Push-T, a robot block-pushing simulation.
Alchaer's first question was how much LeWM is limited by its predictor quality. Swapping it out for perfect simulator physics gave nearly identical results: 78% vs. 77% success on long-horizon tasks. The predictor is already about as good as real physics along trained trajectories—so the bottleneck lies elsewhere.
Out of the box, LeWM solves only 22% of 75-step tasks, versus 89.7% for 25-step ones. Alchaer found four planning interventions—none requiring retraining—that pushed 75-step success to 85%. One single change to how the goal state is scored added 60 percentage points, triple the gain from replacing the predictor with true physics. The research also shows LeWM can be extended to out-of-distribution goals using memory-graph transformations to generate subgoals on the fly.