Stage 1 is trained from the Qwen3-VL base model on constructed video retrieval and temporal grounding supervision. It is intended as the initialization for Stage 2 training or as a released checkpoint for ablations.

Usage

Project Page: <a href="https://mlvlab.github.io/VideoSearch-R1/" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">mlvlab.github.io/VideoSearch-R1
Repository: <a href="https://github.com/mlvlab/VideoSearch-R1" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">mlvlab/VideoSearch-R1

Use with the VideoSearch-R1 codebase:

bash

bash scripts/data_construct/download_preextracted_data.bash activitynet
EVAL_GPUS=0 bash scripts/inference/inference.bash activitynet --checkpoint VideoSearchR1/activitynet-stage1

Citation

bibtex

@inproceedings{lee2026videosearchr1,
  title     = {VideoSearch-R1: Iterative Video Retrieval and Reasoning via Soft Query Refinement},
  author    = {Lee, Seohyun and Choi, Seoung and Ko, Dohwan and Kim, Jongha and Kim, Hyunwoo J.},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}