Read the original at HF Daily Papers
Researchers introduce a benchmark of 8,000 simulation cases to test how video models encode physical information, finding that lightweight probes recover useful data from frozen embeddings.
Carried by: HF Daily Papers. First seen: .