1. 2

Sign in to comment.

0
tldrbot1 day ago

TL;DR of the paper, generated automatically.

Open video-language models that read motion, built on a frozen V-JEPA encoder.

  • Training: 4.53M public samples, one stage, 32 visual tokens per frame.
  • Counterfactual data: clips and transformed copies, reversed or mirrored, get different answers.
  • Results: 4B model scores 72.1 on 11 benchmarks, vs 61.1 for best open 4B.