LaV-CoT: Language-Aware Visual CoT with Multi-Aspect Reward Optimization for Multilingual Text-Centric VQA

Table of Contents

• Prompt Design
• Data
• Training Methods
• Multi-aspect Rewards
• Citation

Prompt

Please check ./prompts for both generator and evaluator prompts.

Data

We design an automatic data curation method that produces scalable, high-quality multilingual CoT annotations through iterative generation, correction, and refinement. All images are resized to 896 * 896.

Language-aware Multi-aspect Rewards

1. Language Reward
2. Count Reward
3. Answer Reward
4. Format Reward

Training Methods

We adopt TRL offical training scripts (https://github.com/huggingface/trl) to do both SFT and GRPO training.

Citation

@misc{huang2025lavcotlanguageawarevisualcot,
      title={LaV-CoT: Language-Aware Visual CoT with Multi-Aspect Reward Optimization for Real-World Multilingual VQA}, 
      author={Jing Huang and Zhiya Tan and Shutao Gong and Fanwei Zeng and Joey Tianyi Zhou and Changtao Miao and Huazhe Tan and Weibin Yao and Jianshu Li},
      year={2025},
      eprint={2509.10026},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.10026}, 
}

For more work, please refer to .