Xiangzheng Zhang

Thinking with Reasoning Skills: Fewer Tokens, More Accuracy

Guangxiang Zhao, Qilong Shi, Xusen Xiao, et al.

VEPO: Variable Entropy Policy Optimization for Low-Resource Language Foundation Models

Chonghan Liu, Yimin Du, Qi An, et al.

This paper introduces VEPO, a new training method that helps AI language models work better with low-resource languages (languages with less training data). The method uses reinforcement learning with built-in quality checks to ensure the model produces properly formatted and grammatically correct outputs, while also dynamically balancing between exact accuracy and natural-sounding responses. Tests show VEPO significantly improves translation quality and efficiency for underrepresented languages.