AI Coders Find Files Faster by Reading Folder Names

When an AI coding assistant is asked to fix a bug, its first problem isn't writing code — it's figuring out which of thousands of files to even look at. Today most tools do this the way a search box does: they look for exact words. The trouble is that the files you actually need often don't contain the words you typed.

A new paper from researcher Manoj Bajaj tested a different idea: let the AI read folder and file names, like a person scanning a filing cabinet. Across 82 real bug-fix tasks pulled from 11 software projects, the name-reading approach surfaced the correct file within its top eight guesses 46.5% of the time. Standard keyword search managed 35.2%, and a single broad text query got just 24.5%. That's roughly a third more correct files found — no code changes required, just better signposting.

So what's the catch? Speed. Reading names took about 9 seconds per task, while keyword search answered in 7 milliseconds — more than a thousand times faster. The study also couldn't prove the AI actually fixed more bugs as a result, only that it found the right files more often. And a more aggressive brute-force approach reached 57.2% accuracy, but at nearly three times the cost in AI calls per task. The author is upfront that the gain might come from something other than folder structure itself.

For anyone who has ever watched an AI assistant confidently edit the wrong file, this matters. Better file-finding means less wasted time, fewer weird mistakes, and AI helpers that understand messy real-world projects. It's not a product yet — it's a signpost for where coding tools are heading.