LeWorldModel tackles JEPA collapse from pixels

This feels less like a flashy benchmark stunt and more like a credible attempt to make JEPA a practical training recipe.

• –The main win is collapse avoidance without the usual scaffolding: no stop-gradient tricks, no EMA teacher, no frozen encoder, just two losses.
• –The compute story is strong: about 15M trainable parameters, single-GPU training in hours, and planning up to 48x faster than foundation-model-based world models.
• –The pure-pixel setup is a meaningful stress test because it removes proprioception and pushes the latent space to carry the planning signal itself.
• –The caveat matters: DINO-WM still has an edge on the visually harder OGBench-Cube task, and LeWM underperforms on Two-Room, so the regularizer is not a universal fix.
• –The most interesting part is the physical-structure signal, especially surprise detection for implausible events, because that suggests the model is learning more than just low-loss predictions.