一句话：提出将覆盖问题变换到局部空间并施加数值方法与线性代数的灰盒模糊测试方法，提升分支覆盖。
论文形式化定义并研究自动测试生成中的覆盖问题：给定一条通向分支点的执行路径及其输入向量，目标是修改输入，使程序仍沿同一路径执行，但分支点取相反真值。标准模糊测试依赖位翻转、随机变异和梯度下降。作者提出借助数值方法与线性代数，将问题变换到所谓局部空间，再在局部空间中应用这些标准技术。在 Test-Comp 2025 基准上的实验表明，该方法在所研究的覆盖问题以及最大分支覆盖测试生成上均优于标准模糊测试技术。
This paper formally defines and tackles a key problem of automated test generation. The coverage problem is specified by a program execution path leading to a certain branching point and an input vector of this execution. The goal is to modify the input vector such that the program running on the modified input follows the same path but the branching point is evaluated to the opposite truth value. Standard fuzzing techniques use bit flips, random fuzzing, and gradient descent to solve this problem. We present an approach that uses numerical methods and linear algebra to transform the problem to so-called local space and applies the mentioned techniques there. An experimental evaluation on the Test-Comp 2025 benchmarks shows some improvements over the standard fuzzing techniques, both for the studied coverage problem and for the problem of test generation for maximal branch coverage.