Jean-Jacques Dubray · Preprints.org 2026 · 2026
DOI: 10.20944/preprints202609.2310.v1
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Joint-embedding predictive architectures (JEPAs) learn a world model by predicting the embedding of the next observation, and plan by minimising the distance to a goal embedding. They are evaluated mostly by probe recovery of the true state and by prediction loss. We audit one against a world built from declared state machines, where the state, transitions, refusals and shortest paths are all known: a two-room navigation task in which every move through the wall is refused and counted. We trained 95 scaled-down LeWorldModels and planned with the cross-entropy method. The position is recovered almost perfectly (probe R² up to 1.00) while planning succeeds on 2 to 26% of tasks, against 100% with the true dynamics; under the latent distance, a larger latent recovers better and plans worse. The latent distance leads the planner away from the door. A critic trained on the frozen latents with privileged labels raised success by 60 points [51, 69] on five unseen Gaussian-start encoders in a declared confirmation, and with it the largest latent appears to plan best. Learned without labels by temporal-difference bootstrapping, the critic recovered 61% of that gain on three isotropic models, but this recipe passed its control on only two of five unseen encoders; it falls short on paths through the door, whose distances the latent carries. Latents trained on narrow goal-directed data fail in the predictor: with a correct cost they still fail until the true dynamics replace it. Prediction loss barely ranks models by planning failure (Spearman 0.34). The representation can be sound while the cost and the predictor fail, and a declared world is a precise instrument for finding out which.
No comments yet — start the discussion below.