Do Not Treat Code as Natural Language: Implications for Repository-Level Code Generation and Beyond

A new research paper argues that treating code as natural language is fundamentally flawed for repository-level AI coding tasks. It introduces 'Hydra', a framework that treats code as structured data, using a structure-aware index and a dependency-aware retriever. On the DevEval and RepoExec benchmarks, Hydra achieved state-of-the-art performance, surpassing the strongest baseline by over 5% in Pass@1 and enabling smaller models to match larger ones.