Vision-SR1: Self-Rewarding Vision-Language Model via Reasoning Decomposition

[📖 Paper]

Models:
🤗 Vision-SR1-7B | 🤗 Vision-SR1-7B-Cold-Start

Datasets:
📊 Vision-SR1-Cold-Start-9K | 📊 Vision-SR1-47K

Training Curves:
📈 Vision-SR1

---

👀 About Vision-SR1

Vision-SR1 is a self-rewarded RL training framework to decompose VLMs' language reasoning into visual perception reasoning and language reasoning. Inspired by the awesome works of e.g. Vision-R1, Visionary-R1, R1-VL, we leverage VLM's self evolving and reasoning ability to Reward Itself.

Because VLMs fuse the vision encoder with the LLM backbone only late in pretraining, they often rely primarily on language reasoning rather than visual perception. Standard RL training tends to recall prior language knowledge for accuracy gains while neglecting vision. External LLM-based perception rewards can help but introduce bias and heavy latency. We instead propose a self-reward framework, enabling the model to provide its own visual and reasoning feedback with no latency.

Besides vision decomposition, We constructed two datasets: Vision-SR1-Cold-9K for SFT and Vision-SR1-47K for RL.

🔍 Dataset

Our training dataset is sourced from 23 sources and evenly split across three main areas -- general visual understanding, science knowledge, multimodal mathematical reasoning.

New Features:

-- Supports Lora Training. Results are not verified. -- Support Qwen3-VL series. However, Qwen3 series format reward is always 0. (Pending debug.) -- Separate advantage computation for final answer accuracy and visual description accuracy.

Requirements

The code base adopted from verl and EasyR1.

Software Requirements

• Python 3.9+
• transformers=4.49.0

Setup

git clone https://github.com/zli12321/Vision-SR1.git
cd Vision-SR1
conda create -n Vision-SR1 python=3.12
bash setup.sh

Training

We support both full fine-tuning and LoRA fine-tuning for two training modes:

• Vision-R1: Standard GRPO with accuracy reward (<think>...\boxed{} format).
• Vision-SR1: Self-reward GRPO with accuracy + self-generated perception reward (<description>...<think>...\boxed{} format).

Full Fine-Tuning

# Vision-SR1 (Self-Reward) full fine-tuning
bash ./vision_sr1/train.sh

# Vision-R1 (standard accuracy) full fine-tuning
bash ./vision_r1/train.sh

• Checkpoints are saved to ./saves/7b_grpo_self_reward/ and ./saves/7b_grpo_accuracy/ respectively.

• To use Qwen3-VL, simply change the model name in train.sh file.

LoRA Fine-Tuning

# Vision-SR1 LoRA fine-tuning
bash ./vision_sr1_lora/train.sh

# Vision-R1 LoRA fine-tuning
bash ./vision_r1_lora/train.sh

LoRA training uses rank=64, lr=1e-5, and only trains the language model layers (vision tower is excluded via exclude_modules: .*visual.*). Checkpoints are saved to ./saves/7b_grpo_self_reward_lora/ and ./saves/7b_grpo_accuracy_lora/ respectively.

Merge Checkpoints

python3 scripts/model_merger.py --local_dir CHECKPOINT_SAVE_DIR/global_step_*/actor

Hardware Requirements

* estimated

Method	Bits	3B	7B
GRPO Full Fine-Tuning	AMP	4/8 or 8x80GB	4/8 or 8x80GB
GRPO LoRA Fine-Tuning	AMP	2/4/8 or 8x32GB	2/4/8 or 8x40GB

Note

Use worker.actor.fsdp.torch_dtype=bf16 and worker.actor.optim.strategy=adamw_bf16 to enable bf16 training with fewer memory.

Evaluation

The evaluation/ folder contains scripts for evaluating checkpoints across multiple benchmarks with automated answer extraction and LLM-based judging.

Folder Structure

evaluation/
├── eval_config.yaml              # Base evaluation config (val_only, greedy decoding)
├── format_prompt/
│   ├── cot_format.jinja           # CoT prompt template (for Vision-R1)
│   └── see_think_format.jinja     # See-Think prompt template (for Vision-SR1)
├── reward_function/
│   └── eval_accuracy.py           # Rule-based accuracy scoring
├── full_rl/
│   ├── eval_vision_r1.sh          # Evaluate full fine-tuned Vision-R1 checkpoints
│   └── eval_vision_sr1.sh         # Evaluate full fine-tuned Vision-SR1 checkpoints
├── lora_rl/
│   ├── eval_vision_r1_lora.sh     # Evaluate LoRA Vision-R1 checkpoints
│   └── eval_vision_sr1_lora.sh    # Evaluate LoRA Vision-SR1 checkpoints
├── llm_judge.py                   # Extract \boxed{} answers, then judge with LLM
└── print_accuracy.py              # Print accuracy table from judgment files

Running Evaluations

Full Fine-Tuning Checkpoints

# Evaluate base model (no checkpoint):
bash evaluation/full_rl/eval_vision_r1.sh

# Evaluate a specific checkpoint:
bash evaluation/full_rl/eval_vision_r1.sh ./saves/7b_grpo_accuracy/global_step_15

# Vision-SR1 variant:
bash evaluation/full_rl/eval_vision_sr1.sh ./saves/7b_grpo_self_reward/global_step_15

LoRA Checkpoints

# LoRA checkpoint (required):
bash evaluation/lora_rl/eval_vision_r1_lora.sh ./saves/7b_grpo_accuracy_lora/global_step_15

# Vision-SR1 LoRA:
bash evaluation/lora_rl/eval_vision_sr1_lora.sh ./saves/7b_grpo_self_reward_lora/global_step_15

Evaluation Pipeline

Each evaluation script runs the following three-stage pipeline automatically:

1. Generate responses: Run the model on each benchmark dataset with greedy decoding (temperature=0), saving all responses to evaluation/responses/.

2. Extract & judge: llm_judge.py extracts the final answer from \boxed{} in each response, then sends the extracted answer and the ground truth to an LLM judge (Qwen2.5-14B-Instruct via vLLM) for comparison. Judgments are saved to evaluation/judgments/.

3. Print accuracy: print_accuracy.py aggregates the judgments and prints a per-dataset accuracy table comparing LLM-judge accuracy with the rule-based score.

Supported Benchmarks

The evaluation scripts include the following datasets (uncomment as needed in the shell scripts):

mmstar, mm-vet, MLLM_test, visnumbench, mmmu_pro_10options, mmmu-pro-vision, hallusionbench, MMMU, MMSI, mathverse, mathvision, mathvista, realWorldQA

Supervised Finetuning (Cold Start)

The supervised finetuning code is adopted from LLaMA-Factory for easy setup.

Download the filtered SFT format data

while ! python download-sft-data.py; do echo "Retrying..."; sleep 5; done

Setup

conda create -n SFT python=3.11
cd LLaMA-Factory-Cold-Start
pip install -e ".[torch,metrics]" --no-build-isolation

pip install --upgrade huggingface_hub
huggingface-cli login

Training

FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/Vision-SR1-Cold-Start.yaml

Troubleshoot

If you still encounter errors after you follow the setup, simply clone the original LLaMA-Factory repo and follow their setup. Download the dataset and place into the LLaMA-Factory data folder. Place the Vision-SR1-Cold-Start.yaml file into the LLaMA-Factory SFT training folder.

Custom Dataset

Please refer to the example datasets to prepare your own dataset.

• Text dataset: https://huggingface.co/datasets/hiyouga/math12k
• Image-text dataset: https://huggingface.co/datasets/hiyouga/geometry3k
• Multi-image-text dataset: https://huggingface.co/datasets/hiyouga/journeybench-multi-image-vqa

Reward Progression in Training

Citation

If you find our works helpful, please cite

@misc{li2025selfrewardingvisionlanguagemodelreasoning,
      title={Self-Rewarding Vision-Language Model via Reasoning Decomposition}, 
      author={Zongxia Li and Wenhao Yu and Chengsong Huang and Rui Liu and Zhenwen Liang and Fuxiao Liu and Jingxi Che and Dian Yu and Jordan Boyd-Graber and Haitao Mi and Dong Yu},
      year={2025},
      eprint={2508.19652},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.19652}, 
}

@article{huang2508self,
  title={Self-evolving reasoning llm from zero data, 2025},
  author={Huang, Chengsong and Yu, Wenhao and Wang, Xiaoyang and Zhang, Hongming and Li, Zongxia and Li, Ruosen and Huang, Jiaxin and Mi, Haitao},
  journal={URL https://arxiv. org/abs/2508.05004}
}

@article{he2025visplay,
  title={Visplay: Self-evolving vision-language models from images},
  author={He, Yicheng and Huang, Chengsong and Li, Zongxia and Huang, Jiaxin and Yang, Yonghui},
  journal={arXiv preprint arXiv:2511.15661},
  year={2025}
}

We recommend to also cite the sourcecode work.

@misc{zheng2025easyr1,
  title        = {EasyR1: An Efficient, Scalable, Multi-Modality RL Training Framework},
  author       = {Yaowei Zheng, Junting Lu, Shenzhi Wang, Zhangchi Feng, Dongdong Kuang, Yuwen Xiong},
  howpublished = {\url{https://github.com/hiyouga/EasyR1}},
  year         = {2025}
}