VRPRM is designed to evaluate intermediate reasoning steps for multimodal problems. The model is intended for visual process reward modeling, reasoning-step scoring, and Best-of-N selection for vision-language model outputs.
The
VRPRM paper trains the model with a two-stage recipe:
This model is not intended to be used as a standalone assistant.
1from transformers import AutoModelForVision2Seq, AutoProcessor
2
3model_id = "YOUR_USERNAME/VRPRM-Qwen2.5VL-7B"
4
5processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
6model = AutoModelForVision2Seq.from_pretrained(
7 model_id,
8 torch_dtype="auto",
9 device_map="auto",
10 trust_remote_code=True,
11)
For the complete inference and evaluation pipeline, use the VRPRM project code.
1@misc{chen2026vrprmprocessrewardmodeling,
2 title={VRPRM: Process Reward Modeling via Visual Reasoning},
3 author={Xinquan Chen and Chongying Yue and Bangwei Liu and Xuhong Wang and Yingchun Wang and Chaochao Lu},
4 year={2026},
5 eprint={2508.03556},
6 archivePrefix={arXiv},
7 primaryClass={cs.LG},
8 url={https://arxiv.org/abs/2508.03556},
9}
