A research team including Sonia Joseph, Quentin Garrido, and colleagues published a study on July 03, 2026, titled "Interpreting Physics in Video World Models." The research investigates how large-scale video encoders represent physical variables, examining whether these models utilize factorized states or task-specific distributed representations. Through methodologies including layerwise probing, subspace geometry, patch-level decoding, and targeted attention ablations, the team characterized the organization of physical information within these models.

The findings identify a distinct structural transition dubbed the "Physics Emergence Zone," an intermediate-depth layer in the architecture where physical variables transition to linear accessibility. Scalar properties such as speed and acceleration are observable in early layers of the models. However, motion direction only becomes accessible at the Physics Emergence Zone. This progression parallels the motion hierarchy observed in the V1 to MT visual processing areas of the primate visual cortex.

The study reveals that motion direction is encoded as a circular high-dimensional population code (a method where neural activity represents variables via patterns across many neurons). Researchers observed that dozens of orthogonal probe dimensions must be steered concurrently to modify the decoded direction, which requires significantly more intervention than the low-dimensional steering used in language models. The evidence refutes the presence of compact, engine-like physical state variables, instead supporting a model of distributed, hierarchically-organized representations that remain effective for physical prediction tasks. The research was published by the International Conference on Machine Learning (ICML) and focuses on the intersection of theoretical machine learning and human-machine intelligence.

A research team including Sonia Joseph, Quentin Garrido, and colleagues published a study on July 03, 2026, titled "Interpreting Physics in Video World Models." The research investigates how large-scale video encoders represent physical variables, examining whether these models utilize factorized states or task-specific distributed representations. Through methodologies including layerwise probing, subspace geometry, patch-level decoding, and targeted attention ablations, the team characterized the organization of physical information within these models.

The findings identify a distinct structural transition dubbed the "Physics Emergence Zone," an intermediate-depth layer in the architecture where physical variables transition to linear accessibility. Scalar properties such as speed and acceleration are observable in early layers of the models. However, motion direction only becomes accessible at the Physics Emergence Zone. This progression parallels the motion hierarchy observed in the V1 to MT visual processing areas of the primate visual cortex.

The study reveals that motion direction is encoded as a circular high-dimensional population code (a method where neural activity represents variables via patterns across many neurons). Researchers observed that dozens of orthogonal probe dimensions must be steered concurrently to modify the decoded direction, which requires significantly more intervention than the low-dimensional steering used in language models. The evidence refutes the presence of compact, engine-like physical state variables, instead supporting a model of distributed, hierarchically-organized representations that remain effective for physical prediction tasks. The research was published by the International Conference on Machine Learning (ICML) and focuses on the intersection of theoretical machine learning and human-machine intelligence.