AI Coding Assistants Learn to Find the Files They Missed
Fewer wrong turns by AI coders means less time fixing their mistakes.
When an AI coding assistant works on a bug, it looks things up in your code — a bit like a researcher pulling files from a filing cabinet. The problem: it searches for whatever looks most similar to the question. Halfway through a task, the assistant has already read the obvious files. So similarity search keeps handing it more of the same thing, while the one missing fact goes unfound. Imagine packing for a trip and grabbing five pairs of socks but no passport.
A research team from the University of California, San Diego describes a fix they call set construction rather than ranking. Instead of scoring files one at a time, the system asks: what does this next decision actually still lack? Three quick AI checks propose a complete set of evidence, hunt for what is missing, and return four to eight intact chunks of real source code inside a 6,144-token budget — roughly 4,600 words, about a short book chapter.
The numbers are modest but real. On 500 frozen test situations drawn from 45 software projects, the method assembled a complete, decision-ready set 73% of the time when allowed five items, and 80.6% with eight. The standard approach (search by similarity, then re-sort) managed 61.4% and 72.4%. A control that ranked purely by resemblance hit 66.6%, showing the gain comes from the strategy, not extra computing. Deleting just one required piece of evidence dropped accuracy by more than 11 points, confirming every piece earns its place.
The honest catch: this is a research paper, not a product you can buy. Its tests use annotated answers, meaning a human labeled which facts each decision truly needed — real-world code rarely arrives that neatly. And even the best configuration still fails roughly one time in five. Still, it points to a clear direction: AI helpers that ask what they are missing, not just what looks familiar.
- AI coding helpers currently search by similarity, so they often return the same fact repeatedly and miss the detail that actually matters.
- The new method builds a complete set of needed evidence instead: 73% success at five items versus 61% for the standard approach.
- On a related test it used a 76% smaller prompt while scoring slightly higher — meaning cheaper, faster AI help.
- Catch: it relies on human-labeled 'correct answers' in testing, and still fails about one in five times.
Why It Matters
Better-fed AI coders mean fewer wrong fixes, faster software, and less time cleaning up machine mistakes.