Causal-JEPA (C-JEPA)

Key Claims

• Extends the masked JEPA framework to object-centric representations
• Object-level masking induces counterfactual-like effects in the latent space
• Aims to instill causal structure into learned world models rather than leaving causality implicit

Why This Matters

LeCun's critique of LLMs includes that they're good at correlation, poor at intervention. A world model that can answer "what if this object weren't here?" is meaningfully more than a video predictor. C-JEPA is one of the first serious attempts to bake intervention into JEPA.

Notes

First-pass stub from search results. Flag for deeper read if causal world models become a distinct frontier.

Source: Causal-JEPA