Ce Zhang, Ziyang Wang, Yulu Pan +6 more
Answering a question about a long video means finding the few seconds that actually contain the answer. Recent systems do this by repeatedly cropping: look at a stretch, decide the answer is somewhere in the second half, crop to that, repeat. It narrows quickly, which is the appeal.
The flaw is that cropping only ever goes one direction. Once the agent has committed to a stretch of video, it has no way to climb back out and reconsider, so an early wrong turn is permanent. The paper describes this as converging prematurely with no way to recover.
VideoTreeSearch reframes the task as searching a tree rather than narrowing a window. The tree is built from visual scene boundaries, so its divisions follow where the video actually changes instead of arbitrary time slices, and the agent can move back up as well as down. It is the difference between tearing pages out of a book as you narrow your search and simply keeping a finger in the earlier chapter in case you need to go back.
Grounded long-video question answering (Grounded LVQA) requires answering a question about a long video while localizing the short evidence interval that supports the answer. Recent agentic methods frame this task as multi-turn exploration with a single crop_video(start, end) action, which supports coarse-to-fine narrowing but provides no primitive for fine-to-coarse backtracking. As a result, these agents typically converge prematurely and cannot recover from an early…
PRISA: Proactive Infrastructure LiDAR Framework for Intersection Safety Assessment
arXiv (cs.CV) · July 17, 2026CLIFE: Camera-LiDAR Fusion Framework for Edge-Deployable Roadside VRU Perception
arXiv (cs.CV) · July 17, 2026VTLoc: Learning-based Tactile Contact Localization in Visual Point Clouds
arXiv (cs.LG) · July 17, 2026Learning Standard Model structure from LHC data with Riemannian flow matching