In this work, we introduce Chain of Step reasoning for vision-language models, enabling assessing reasoning step quality accurately and leading to effective reinforcement learning and inference-time scaling with fine-grained rewards. Experimental results across multiple benchmarks demonstrate the effectiveness of CoS. More importantly, we conduct extensive empirical analysis and ablations to unveil CoS’s appealing properties. We hope this paper offers insights into more complex multi-modal reasoning.

ShareGPT-Step-300K

Note: You can directly use our SFT dataset (special tokens have been added) through the following link, or you can assess the raw step data to customize your SFT dataset. For customization, you can modify get_sft_json.py to get your SFT data accordingly.

	Description	Links
ShareGPT-Step-300K.jsonl	The SFT Jsonl	🤗 HF link
images.zip	image files	🤗 HF link
raw_jsonl.zip	raw step jsonl file for customization	🤗 HF link

PRM & Data

Note: You can directly use our train jsonl file to train the PRM (special tokens have been added with a fixed format) through the following link, or you can assess the raw data to customize your dataset. For customization, you can modify get_prm_json.py to get your data accordingly.

	Description	Links
CoS-PRM	The PRM model	🤗 HF link
prm_data_raw.json	raw prm data	🤗 HF link
prm_data_train.jsonl	prm training jsonl	🤗 HF link

Checkpoints

	Description	Links
CoS-SFT	The SFT model	🤗 HF link
CoS	The RL model	🤗 HF link

ToDo List

• SFT Dataset
• PRM & Dataset
• Training & Inference code
• Checkpoints

License

Apache License 2.0

Citation

@article{chen2025unveiling,
  title={Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards},
  author={Chen, Honghao and Lou, Xingzhou and Feng, Xiaokun and Huang, Kaiqi and Wang, Xinlong},
  journal={arXiv preprint arXiv:2509.19003},
  year={2025}
}

Acknowledgement

We thank the repositories for their excellent work: InternVL, LLaVa-NeXt, TAP