What's this about?
A paper tackling why multimodal LLMs are good at recognizing what's in an image/video (passive perception) but weak at reasoning over how things change across time. It reconstructs a hidden "state" to support multi-step temporal reasoning.
What exactly is "temporal reasoning"?
Temporal reasoning is understanding sequence and change over time — what happened before/after, cause and effect across frames — not just what a single frame shows. It's what you need to answer "why did the ball fall?" from a video rather than "there is a ball."
Why it matters
Video understanding and embodied agents both need this; it's a known weak spot for current multimodal models.