CogVideoX

Zhipu AI / THUDM

Structures found in this family (2)

Anisotropy, Linear Separability

By model (3)

CogVideoX-2B · 2B

Linear Separability

CogVideoX-5B

Anisotropy

LTX-Video (diffusion-based video generator)

Linear Separability

Observations (3)

Video-DiT massive activations follow a token-position hierarchy

Inverting video diffusion transformers makes physical plausibility linearly decodable

Video foundation models vary in physics decodability, V-JEPA strongest

Papers

Steering Video Diffusion Transformers with Massive Activations (2026), The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show (2026), Do Video Foundation Models Understand Intuitive Physics? A Layerwise Probing Analysis (2026)