VeriScale: Adversarial Test-Suite Scaling for Verifiable Code Generation
VeriScale:对抗性测试套件缩放用于可验证代码生成
机构 * School of Mathematical Sciences, Shanghai Jiao Tong University(上海交通大学数学科学学院) ; School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen(香港中文大学(深圳)科学与工程学院) ; School of Mathematics, Jilin University(吉林大学数学学院) ; School of Mathematical Sciences, Tongji University(同济大学数学科学学院) ; Zhiyuan College, Shanghai Jiao Tong University(上海交通大学紫阳学院) ; School of Future Technology, South China University of Technology(华南理工大学未来技术学院) ; Institute of Natural Sciences, Shanghai Jiao Tong University(上海交通大学自然科学研究院) ; MOE-LSC, CMA-Shanghai, Shanghai Jiao Tong University(上海交通大学MOE-LSC、CMA-上海)
AI总结 本文提出VeriScale框架,通过对抗性实现扩展和缩减测试套件,提升代码生成的可验证性,实验表明VerinaPlus显著暴露了模型弱点,而VerinaLite在低成本下保持判别能力。