The authors paired each video with a version whose frames were reversed, preserving visual content and text while changing temporal order.
Video models detect temporal order internally, then let the signal fade
A training-free intervention preserves temporal information that emerges in intermediate layers but weakens before prediction.
Top University
Youngwoo Shin · Yusung Ro · Minseo Kim · Junmo Kim
Korea Advanced Institute of Science and Technology (KAIST)
Research Digest··3 min read
Shin et al.
Why this paper
From Korea Advanced Institute of Science and Technology (KAIST)
In one line
Temporal information in VideoLLMs peaks in intermediate layers and fades before output, and reinjecting it improves temporal reasoning without training.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§