Check out Cambrian-P, a must-read if you work on video/3D MLLMs or spatial understanding.
Key takeaway: camera pose, which naturally describe the relationships between isolated frames, turns out to be highly effective for enhancing video understanding and spatial intelligence!