• October 17, 2025. We have released the inference code.
• September 29, 2025. We have released the training code.
• September 28, 2025. We have released the PeBR-R1 model weights and the PeBR-R1 dataset on Hugging Face.
• September 16, 2025. Our paper was released on arXiv.

Reinforcement learning (RL) has proven highly effective in eliciting the reasoning capabilities of large language models (LLMs). Inspired by this success, recent studies have explored applying similar techniques to vision-language models (VLMs), aiming to enhance their reasoning performance.

However, directly transplanting RL methods from LLMs to VLMs is suboptimal, as the tasks faced by VLMs are inherently more complex. Specifically, VLMs must first accurately perceive and understand visual inputs before reasoning can be effectively performed.

To address this challenge, we propose a two-stage reinforcement learning framework designed to jointly enhance both the perceptual and reasoning capabilities of VLMs. To mitigate the vanishing advantage issue commonly observed in RL training, we first perform dataset-level sampling to selectively strengthen specific capabilities using distinct data sources.

During training:

• Stage 1 (Perception RL): improves the model’s visual perception through coarse- and fine-grained visual understanding.
• Stage 2 (Reasoning RL): targets the enhancement of reasoning abilities.

After the proposed two-stage reinforcement learning process, we obtain PeBR-R1, a vision-language model with significantly enhanced perceptual and reasoning capabilities.

Experimental results on seven benchmark datasets demonstrate the effectiveness of our approach and validate the superior performance of PeBR-R1 across diverse visual reasoning tasks.

conda create -n pebrr1 python==3.10
conda activate pebrr1
git clone https://github.com/cythu/PeBR-R1.git
cd PeBR-R1
pip install -r environment.txt
pip install -e .

To prepare the dataset for SFT warm-up:

1. Download and filter Mulberry-SFT Get the Mulberry-SFT dataset Run the filtering script to remove reflective reasoning samples and any images with width or height smaller than 28 pixels:

python dataset/filter.py --input dataset/mulberry_sft.json --output dataset/pebr_sft.json --image_root dataset/mulberry_images

2. Merge with PeBR-R1 SFT data Download the PeBR-R1 sft dataset and merge its images and JSON files with the filtered Mulberry-SFT data.

3. Train with LLaMA-Factory Register the merged dataset in LLaMA-Factory by updating dataset_info.json.

Example entry:

"pebr_r1_warm_up": {
    "file_name": "./pebr_sft.json",
    "formatting": "sharegpt",
    "columns": {
      "messages": "messages",
      "images": "images"
    },
    "tags": {
      "role_tag": "role",
      "content_tag": "content",
      "user_tag": "user",
      "assistant_tag": "assistant"
    }
  }

4.Training configuration Use the config file provided in ./docs/qwen2_5vl_full_sft.yaml.

Download the GRPO dataset and place the files into the following locations:

• Images → ./dataset/grpo/pebr_grpo_images
• JSON file → ./dataset/grpo/pebr_grpo_dataset.json

1. Run the sampling script to generate rollout data—each question produces 8 outputs:

bash examples/pebr_sample.sh

2. Split the dataset into easy_case, medium_case, and hard_case based on output accuracy:

python _rollout_outputs/sample_dataset.py

In this stage, we focus on enhancing the perception ability of the VLM.

1. Start Stage 1 RL training:

bash examples/pebr_stage1.sh

2. Merge the resulting model:

python scripts/model_merger.py --local_dir checkpoints/easy_r1/stage1_rl/global_step_1/actor

In this stage, we focus on strengthening the reasoning ability of the VLM.

1. Start Stage 2 RL training:

bash examples/pebr_stage2.sh

2. Merge the resulting model:

python scripts/model_merger.py --local_dir checkpoints/easy_r1/stage2_rl/global_step_1/actor

After downloading the model from Hugging Face, you can run inference with the following command:

python inference.py

Example inputs (image and question) can be customized inside inference.py.

• LLaMA-Factory: We use this framework for SFT warm-up training.
• EasyR1: We build our RL training pipeline on top of this project.
• VLMEvalKit: We adopt this toolkit for evaluating vision-language models.

We sincerely thank the authors and contributors of these excellent open-source projects for enabling our work.

If you find PeBR-R1 useful for your research, please consider citing our work:

@article{chen2025perception,
  title={Perception Before Reasoning: Two-Stage Reinforcement Learning for Visual Reasoning in Vision-Language Models},
  author={Chen, Yan and Li, Long and Xi, Teng and Zeng, Long and Wang, Jingdong},
  journal={arXiv preprint arXiv:2509.13031},
  year={2025}
}