一句话：提出类型批处理程序约简：用机器学习推荐优先约简的程序批次，并支持多批次同时约简以加速跨语言程序缩减。
程序约简（delta debugging）在保持给定性质的前提下为程序寻找更小、更易理解的版本。领域无关的约简器可跨语言复用，却常常需要数小时才能完成，严重影响开发效率与可扩展性。本文提出类型批处理程序约简（type batched program reduction）：利用机器学习在约简过程的特定阶段推荐最可能带来收益的程序片段（即批次），并进一步扩展为一次性联合约简多个片段。实验在来自多种编程语言的一组大型程序上，以约简时间为指标与当前最先进的约简器进行比较，结果表明本方法在约简时间上优于现有工作。该工作为机器学习引导的程序约简奠定了基础。
Given a program with a property of interest, program reduction searches for a smaller program that preserves the property and is easier to understand. Domain agnostic program reducers can reduce programs of multiple languages without extra domain knowledge. Despite their reusability, they may still take hours to run, hindering productivity and scalability. This paper proposes type batched program reduction, which uses machine learning to suggest portions of a program, or batches, that are most likely to be advantageous to reduce at a particular point in the reduction. We also extend this to jointly reduce multiple portions of a program at once, improving the performance further. Suggesting an appropriate order for removing batches from a program along with their potential simultaneous removal enables our reducer to outperform the state of the art reducers in reduction time over a set of large programs from multiple programming languages. This work lays foundations for further improvements in ML guided program reduction.