Researchers test whether AI code repair works better with exact function locations versus entire files

Computer scientists tested how precisely you need to pinpoint a bug's location before feeding it to an AI that fixes code automatically—and found that function-level precision (the specific function, not just the file) works best. This matters because many automated code repair systems separate bug-finding from bug-fixing into two steps, and the granularity of the first step directly affects whether the second step succeeds.

This is a proof-of-concept study that isolates a specific variable—how much precision in bug localization actually helps AI repair systems work—that hadn't been systematically tested at real repository scale before. The authors aren't claiming to beat state-of-the-art systems; they're asking a simpler structural question: 'Does the AI repair engine care whether we tell it the exact function versus the whole file versus just the line number?' The answer appears to be yes, and it matters differently depending on the specific bug. This kind of methodological isolation—testing one thing in controlled conditions while holding everything else constant—is how you move past marketing claims to understanding what actually moves the needle.