Reflect-R1

Model checkpoints for Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding.

• Paper: https://arxiv.org/abs/2606.27922
• Code: https://github.com/ShuimuChen-hyq/Reflect-R1
• Data: https://huggingface.co/datasets/CSDDSFSFSAFSAF/Reflect-R1-data

Checkpoints

Reflect-R1-SFT-6000/       Cold-start SFT checkpoint.
Reflect-R1-GRPO-Final/     Final SD-GRPO checkpoint.

Both checkpoints are based on Qwen2.5-VL-7B and include sharded safetensors weights together with the corresponding tokenizer and processor configuration files.

Citation

@article{chen2026reflectr1,
  title   = {Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding},
  author  = {Shuimu Chen and Yuteng Chen and Yuanshen Guan and Zebang Cheng and Zeyu Zhang and Shengqian Qin and Bin Xia and Jiaran Li and Wenming Yang and Fei Ma},
  journal = {arXiv preprint arXiv:2606.27922},
  year    = {2026}
}