[ACM MM 2026] Remember-R1: Mitigating Long-Context Visual Forgetting through Reinforcement Learning

• [2026.08.04] Code released; Models and Datasets released on Hugging Face.
• [2026.08.04] Paper released on arXiv.
• [2026.07.10] 🎉 Remember-R1 was accepted by ACM MM 2026!

During training, Remember-R1 augments the answer correctness reward with three process-level rewards defined on the target reasoning trajectory:

• Visual Vocabulary Reward encourages explicit expression of visual evidence.
• Visual Memory Reward encourages persistence of visual dependence into later reasoning steps.
• Visual Key-Region Reward encourages attention to question-relevant regions.

git clone https://github.com/Ch921-cell/Remember-R1.git
cd Remember-R1

conda create -n remember-r1 python=3.10 -y
conda activate remember-r1
pip install -r requirements.txt
pip install -e src/open-r1-multimodal

The default launcher uses DeepSpeed ZeRO-3 and FlashAttention 2. Install versions compatible with your CUDA and PyTorch environment.

Download the training datasets from the dataset release page and place them in the dataset/ directory.

Configure paths and launch training:

export MODEL_PATH="/path/to/model"
export DATA_PATHS="/path/to/train.jsonl"
export IMAGE_FOLDERS="/path/to/dataset"

bash run_scripts/run_grpo_remember.sh

We evaluate Remember-R1 using VLMEvalKit.

1. Install VLMEvalKit:

   git clone https://github.com/open-compass/VLMEvalKit.git
   cd VLMEvalKit
   pip install -e .

2. Prepare the benchmark datasets according to the VLMEvalKit documentation.

3. Configure the model checkpoint and evaluation settings in VLMEvalKit, then run:

   python run.py --model <model_name_or_path> --data <dataset_name>

If you find Remember-R1 useful in your research, please cite our paper:

@inproceedings{chen2026rememberr1,
    title={Remember-R1: Mitigating Long-Context Visual Forgetting through Reinforcement Learning}, 
    author={Jianmin Chen and Jiaqi Tang and Wei Wei and Xiaogang Xu and Jiafei Wu and Zhe Liu and Qianzhou Wang and Yingying Yan and Botong Geng and Yuyang Xia and Lei Zhang and Qifeng Chen},
    booktitle={Proceedings of the 34th ACM International Conference on Multimedia},
    year={2026}
}

We thank the authors of VLM-R1, VLMEvalkit and VAPO for their open-source contributions.