Generating Test Cases by Recomposing Existing Developer Tests

Abstract

Automated test generation can reach large portions of code, yet it often fails to expose deep faults because it struggles to construct semantically valid, constraint-satisfying object states that actually trigger failures. We present BOOSTER, a developer-test amplification technique that increases fault exposure by harvesting and globally recomposing valid input ingredients from developer-written tests. BOOSTER extracts object-defining statements and primitive/string values into type-indexed pools, augments missing receiver and parameter types via recursive dependency synthesis, and instantiates masked test skeletons with combinatorial selection to efficiently explore input combinations under a fixed budget. By leveraging the architectural knowledge embedded in developer tests, BOOSTER overcomes the limitations of from-scratch generation and addresses the structural constraints of prior amplification techniques that lack global ingredient recomposition. We evaluate BOOSTER on 790 real-world bugs from Defects4J (v2.0.1) across 17 open-source Java projects. In the regression scenario, BOOSTER outperforms auto- mated generation tools, detecting 298 bugs compared to 270 by EVOSUITE and 177 by RANDOOP, and it significantly surpasses state-of-the-art amplification techniques, detecting 235 bugs compared to 113 by EVO-AMP and 25 by DSPOT. In the non-regression scenario, BOOSTER demonstrates dominant performance, detecting 160 bugs compared to 87 by EVOSUITE and 77 by RANDOOP.