-
1
Identifiability Without Gaussianity: Symbolic World Models and Near-Infinite Temporal Consistency
Responding to Klindt, LeCun & Balestriero (arXiv:2605.26379), who proved that Joint-Embedding Predictive Architectures (JEPAs) achieve linear identifiability of the world's true latents iff the latent dynamics are Gaussian and stationary, this paper shows the Gaussian boundary is an artifact of statistical alignment, not a property of world models in general. The authors introduce the Physics-Grou...
-
2
MoVerse: Real-Time Video World Modeling with Panoramic Gaussian Scaffold
MoVerse turns a single narrow-field-of-view image into an interactively navigable 360° scene at 8 FPS on a single RTX 4090, by separating world construction from observation rendering. A topology-aware diffusion model expands the narrow-FoV input into a gravity-aligned 360° panorama; a panoramic geometry-aware residual predictor lifts the panorama into a persistent 3D Gaussian scaffold; a Gaussian...
-
3
Diffusion Transformer World-Action Model for AV Scene Prediction
This paper argues that the autonomous-driving community's standard distortion metrics (cosine similarity, SSIM) actively mislead latent world-model evaluation by rewarding a blurry regression mean over a realistic prediction. A compact latent DiT world-action model, given a present front-camera latent and an ego-action sequence, predicts future scene latents that a frozen decoder renders to 256×25...
-
4
EPM-JEPA: Operator-Side Experience Modulation in JEPA-Family World Models
EPM-JEPA compares two mechanisms for injecting accumulated experience into a JEPA predictor when test-time dynamics diverge from training: operand-side injection (residual addition to hidden state, "EI-JEPA") and operator-side modulation (low-rank weight deltas via LoRA, "EPM-JEPA"). On a pre-registered Moving-MNIST gravity-shift benchmark, the operator-side variant is a null result against operan...