When Does Context Help? A Controlled Study ofLLM-Based Bug Fixing
Large language models (LLMs) have shown promise for automated program repair, but it remains unclear which debugging signals are most useful and when additional context becomes distracting, costly, or ineffective. We present a controlled empirical study of LLM-based bug fixing on FIXEVAL, comparing three model families...