Not All Invariants Are Equal: Curating Training Data to Accelerate Program Verification with SLMs
并非所有不变式都同等重要:利用SLM通过精选训练数据加速程序验证
机构 * Hebrew University of Jerusalem, Israel(特拉维夫大学) ; Amherst College, USA(阿默斯特学院) ; VMware Research by Broadcom, USA(Broadcom VMware 研究)
AI总结 提出Wonda数据筛选流水线,通过AST归一化和LLM语义重写从原始验证器输出中提取高质量训练不变式,微调小语言模型(SLM)后,在多个模型上使不变式正确性和加速率翻倍甚至三倍,小模型性能媲美大模型。
Comments A preprint of the ICML 2026 paper with the same title