suyong-3/ReQuest
Viewer • Updated • 920k • 98
Question-aware frame selector checkpoints for the ECCV 2026 paper
"ReQuest: Rethinking-based Question-Aware Frame Selection for Long-Form Video QA"
[Paper] [Code]
ReQuest is a plug-and-play frame selection framework for long-form video QA. When the MLLM (LLaVA-Video) is uncertain about its first answer, the selector scores every 1 fps frame against the question, and the MLLM answers again from the selected frames. See the code repository for details.
| File | Benchmark | Epoch | tau0 |
Accuracy |
|---|---|---|---|---|
request_selector_videomme.pth |
Video-MME (w/o subtitles) | 20 | 0.55 | 65.6 |
request_selector_longvideobench.pth |
LongVideoBench (val) | 2 | 0.30 | 60.1 |
request_selector_mlvu.pth |
MLVU (dev), m-avg | 1 | 0.45 | 71.7 |
Accuracy is that of LLaVA-Video-7B-Qwen2 with ReQuest, as reported in the paper. All three checkpoints are
epochs of one training run on the ReQuest pseudo-labels.
SHA256SUMS lists the checksums of the files.
git clone https://github.com/ailab-kyunghee/ReQuest && cd ReQuest
huggingface-cli download suyong-3/ReQuest --include "request_selector_*.pth" --local-dir checkpoints
bash scripts/eval.sh videomme # or longvideobench, mlvu
@inproceedings{kim2026request,
title = {{ReQuest}: Rethinking-Based Question-Aware Frame
Selection for Long-Form Video {QA}},
author = {Kim, Minkuk and Yun, Suyong and Kim, Young Tae and
Moon, Jinyoung and Choi, Jinwoo and Kim, Seong Tae},
booktitle = {Computer Vision -- ECCV 2026},
series = {Lecture Notes in Computer Science},
volume = {17052},
pages = {386--404},
publisher = {Springer},
address = {Cham},
year = {2026},
doi = {10.1007/978-3-032-37235-2_23}
}