SVBench is the first benchmark that checks whether video generation models can show realistic social behavior, not just pretty pictures.
Robots often act like goldfish with short memories; HiF-VLA fixes this by letting them use motion to remember the past and predict the future.