Hello! Welcome to the repository for DentalGPT! DentalGPT is the first specialized dental Multimodal Large Language Model (MLLM) equipped with advanced complex reasoning capabilities. While general MLLMs struggle to capture fine-grained dental visual details, DentalGPT leverages high-quality domain knowledge injection and reinforcement learning to interpret complex reasoning in dentistry.

Despite its compact 7B parameter scale, DentalGPT achieves superior performance in disease classification and dental VQA tasks, outperforming many state-of-the-art models with over 100B parameters (such as GPT-5 and Gemini-2.5-Pro).

• Massive Scale: Constructed the largest annotated dental multimodal dataset to date with over 120k dental images.
• 2-Stage Training: A novel pipeline involving Multimodal Understanding Enhancement and Reinforcement Learning (RL).
• Advanced Reasoning: Integration of the Group Relative Policy Optimization (GRPO) algorithm to incentivize long chain-of-thought (CoT) reasoning for precise diagnosis in dentistry.

DentalGPT is developed on top of the Qwen2.5-VL-7B-Instruct backbone through a structured 2-stage process:

This stage focuses on injecting high-quality dental domain knowledge:

• Image Captioning: Alignment of visual features with professional dental terminology.
• Instruction Tuning: Enhancing performance on downstream tasks through expert-verified QA pairs.
• Knowledge Density: Training on data with higher knowledge density and professional quality compared to standard GPT-distilled datasets.

We apply the Group Relative Policy Optimization (GRPO) algorithm to strengthen diagnostic logic:

• Thinking Mode: The model is trained to use <think> tags for internal reasoning and reflection before providing a final <answer>.
• Iterative Refinement: Encourages the model to self-correct intermediate counting or identification errors, leading to higher accuracy in complex dental tasks.

DentalGPT demonstrates superior performance across all evaluated benchmarks, establishing it as a leading multimodal foundation model for dental image understanding.

1. Expert-Level Performance with 7B Efficiency: Despite its compact size, DentalGPT consistently outperforms general-purpose models with over larger parameters (such as GPT-5 and Gemini-2.5-Pro) in dental-specific tasks. This highlights the effectiveness of domain-specialized training.
2. Robust Generalization: The model achieves significant gains across diverse modalities, including professional panoramic X-rays and "in-the-wild" intraoral photos taken by patients. It shows a massive improvement over its backbone model, particularly in identifying complex conditions like periapical lesions and impacted teeth.

We are from:

• The Chinese University of Hong Kong, Shenzhen 香港中文大学（深圳）
• Shenzhen Stomatology Hospital (Pingshan) of Southern Medical University 南方医科大学深圳口腔医院（坪山）
• Faculty of Dentistry, The University of Hong Kong 港大牙医学院
• Freedom AI 深圳自由动脉科技有限公司

特别鸣谢智谱华章科技提供支持。

@article{cai2025dentalgpt,
  title={DentalGPT: Incentivizing Multimodal Complex Reasoning in Dentistry},
  author={Cai, Zhenyang and Zhang, Jiaming and Zhao, Junjie and Zeng, Ziyi and Li, Yanchao and Liang, Jingyi and Chen, Junying and Yang, Yunjin and You, Jiajun and Deng, Shuzhi and Wang, Tongfei and Chen, Wanting and Hao, Chunxiu and Xie, Ruiqi and Wen, Zhenwei and Feng, Xiangyi and Ting, Zou and Lin, Jin Zou and Li, Jianquan and Yu, Guangjun and Chen, Liangyi and Wang, Junwen and Jiang, Shan and Wang, Benyou},
  journal={arXiv preprint arXiv:2512.11558},
  year={2025}
}