Interpreting Physics in Video World Models

A long-standing question in physical reasoning is whether video-based models need to rely on factorized representations of physical variables in order to make physically accurate predictions, or whether they can implicitly represent such variables in a task-specific, distributed manner. While modern video world models achieve strong performance on intuitive physics benchmarks, it remains unclear which of these representational regimes they implement internally. Here, we present the first interpretability study to directly examine physical representations inside large-scale video encoders. Using layerwise probing, subspace geometry, patch-level decoding, and targeted attention ablations, we characterize where physical information becomes accessible and how it is organized within encoder-based video transformers. Across architectures, we identify a sharp intermediate-depth transition -- which we call the Physics Emergence Zone -- at which physical variables become accessible. Physics-related representations peak shortly after this transition and degrade toward the output layers. Decomposing motion into explicit variables, we find that scalar quantities such as speed and acceleration are available from early layers onwards, whereas motion direction becomes accessible only at the Physics Emergence Zone. Notably, we find that direction is encoded through a high-dimensional population structure with circular geometry, requiring coordinated multi-feature intervention to control. These findings suggest that modern video models do not use factorized representations of physical variables like a classical physics engine. Instead, they use a distributed representation that is nonetheless sufficient for making physical predictions.

Self-Supervised Learning from Structural Invariance

Joint-embedding self-supervised learning (SSL), the key paradigm for unsupervised representation learning from visual data, learns from invariances between semantically-related data pairs. We study the one-to-many mapping problem in SSL, where each datum may be mapped to multiple valid targets. This arises when data pairs come from naturally occurring generative processes, e.g., successive video frames. We show that existing methods struggle to flexibly capture this conditional uncertainty. As a remedy, we introduce a latent variable to account for this uncertainty and derive a variational lower bound on the mutual information between paired embeddings. Our derivation yields a simple regularization term for standard SSL objectives. The resulting method, which we call AdaSSL, applies to both contrastive and distillation-based SSL objectives, and we empirically show its versatility in causal representation learning, fine-grained image understanding, and world modeling on videos.

seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models

Current self-supervised algorithms mostly rely on transformations such as data augmentation and masking to learn visual representations. This is achieved by inducing invariance or equivariance with respect to these transformations after encoding two views of an image. This dominant two-view paradigm can limit the flexibility of learned representations for downstream adaptation by creating performance trade-offs between invariance-related tasks such as image classification and more fine-grained equivariance-related tasks. In this work, we introduce \emph{seq-JEPA}, a world modeling paradigm based on joint-embedding predictive architecture that leverages architectural inductive biases to resolve this trade-off. Without requiring an additional equivariance predictor or loss term, seq-JEPA simultaneously learns two architecturally segregated representations: one equivariant to the specified transformations and another invariant to them and suited for tasks such as classification. To do so, our model processes a short sequence of different views (observations) of an input image. Each encoded view is concatenated with embeddings corresponding to the relative transformation (action) producing the next observation in the sequence. A transformer encoder outputs an aggregate representation of this sequence, which is subsequently conditioned on the action leading to the next observation to predict its representation. Empirically, seq-JEPA achieves strong performance on equivariant benchmarks and image classification without sacrificing one for the other. Additionally, our framework excels at tasks that inherently require aggregating a sequence of observations, such as path integration across actions and predictive learning across eye movements.