Athena Combines Program Dependence Graphs With Learned Code Representations for Change Impact Analysis, on a New 25-Project Benchmark
Impact analysis has historically leaned on coupling metrics that need execution traces or long co-change histories, both brittle and expensive. Athena (arXiv 2607.23355) instead fuses a system's dependence graph with a conceptual-coupling approach built on deep representation learning for code, requiring neither change history nor execution information. Because prior IA benchmarks covered fewer than ten projects and suffered from tangled commits, the authors built Alexandria from 25 open-source projects using fine-grained bug-fix commit information; on it, the best Athena configuration reaches mRR 60.32%, mAP 35.19%, and HIT@10 81.48%, outperforming a simpler baseline by 10.34%, 9.55%, and 11.68% with statistical significance.
↳ Follow the thread