Separating genuine predictability from silent collapse in a world model's JEPA surprise signals.
Making the counterfactual difference between paired worlds the primary training objective for world models.
Evidence that systematic generalization tracks competence rather than clean factorization of representations.
An anchored, causally-validated account of per-constraint structure in energy-based models, and why it governs reasoning.
Measuring recursive belief depth in language models: how many levels of "I think that you think" they hold, and where it stops.