What is it about?
A long video can contain thousands of frames, but the answer to a question may depend on a brief moment. Our method, ReCrossVLLM, helps video AI work at two levels. First, it scans many frames with a compact representation to find a relevant clip. Then, it examines that clip in greater visual detail. If the clip does not provide the right evidence, the system sends feedback and searches again. During training, it also learns which clip selections better support answering questions. Tests on three video understanding benchmarks and two event-localization datasets show improvements in answering questions and identifying when relevant events occur.
Featured Image
Photo by Hossein Nasr on Unsplash
Why is it important?
Understanding long videos requires both the overall story and small visual details. Compressing every frame can hide useful clues, while viewing only a few frames can miss relevant events. This research offers a way to balance those needs and revisit an unhelpful initial choice. With the same LLaVA-Video backbone, ReCrossVLLM improves results across all three tested video understanding benchmarks. The approach could support more useful video search and question answering, although its reported benefits come from benchmark experiments.
Perspectives
For me, the central idea is that understanding a long video should include the ability to change where the model looks. Finding a plausible clip is only a first step. Combining a broad view, close inspection and feedback gives video AI a practical way to reconsider its evidence before answering.
Wei Feng
Tsinghua University
Read the Original
This page is a summary of: Reflective Cross-Granularity Grounding with Preference Optimization for Long Video Understanding, June 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3805622.3810717.
You can read the full text:
Resources
Contributors
The following have contributed to this page
