RankVideo is a video-native reasoning reranker for text-to-video retrieval, fine-tuned from
Qwen3-VL-8B-Instruct.
The model explicitly reasons over query-video pairs using video content to assess relevance. It was introduced in the paper
RANKVIDEO: Reasoning Reranking for Text-to-Video Retrieval.
This model was trained using the
MultiVENT 2.0 dataset and
RankVideo-Dataset.
1from rankvideo import VLMReranker
2
3reranker = VLMReranker(model_path="hltcoe/RankVideo")
4
5# Score query-video pairs for relevance
6scores = reranker.score_batch(
7 queries=["person playing guitar"],
8 video_paths=["/path/to/video.mp4"],
9)
10
11print(f"Relevance score: {scores[0]['logit_delta_yes_minus_no']:.3f}")
1@misc{skow2026rankvideoreasoningrerankingtexttovideo,
2 title={RANKVIDEO: Reasoning Reranking for Text-to-Video Retrieval},
3 author={Tyler Skow and Alexander Martin and Benjamin Van Durme and Rama Chellappa and Reno Kriz},
4 year={2026},
5 eprint={2602.02444},
6 archivePrefix={arXiv},
7 primaryClass={cs.IR},
8 url={https://arxiv.org/abs/2602.02444},
9}
