PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards

PreciseCoder. Faithful and Generalizable Debugging Beyond Unit-Test Rewards. By Yunkai Zhan and seven others, from the University of Southern California, the University of Chicago, and two other institutions. Published on the thirteenth of October. No code repository is linked to it.

The common recipe for post-training code models is reinforcement learning from unit tests. But a unit test asks only one question: does the final program pass? The authors analyze what that reward teaches a model when it debugs. Pass rates go up. Alongside that, the model learns to edit more than it needs to, and to hallucinate bugs in code that is already correct. Nothing in a pass or fail verdict distinguishes a one-line fix from a sweeping rewrite that happens to pass, so the reward cannot prefer the first. And a model that rewrites correct lines is changing code nobody asked it to touch.

PreciseCoder is a training algorithm that keeps the unit-test reward and adds a second signal, one that rewards edits which are semantically verified and localized. That signal correlates with smaller edits and with higher pass rates.

What the reward adds is an incentive to change only what is wrong, and for the change to be checked as meaningful, on top of the incentive to make the tests pass. The pass signal stays. It simply stops being the only thing the model is paid for.

Results are reported at four billion, nine billion, and twenty-seven billion parameters. Unit-test-only training leaves models editing between seven and fifteen lines on average. PreciseCoder brings that down to three, at a pass rate on par. Trained only on single-line bugs, it generalizes to repository-level program repair, resolving more issues with smaller patches. That is a large step in scope, from one faulty line to issues in whole repositories.

The precision also transfers from weaker models to stronger ones. Given the reasoning trace of the nine billion parameter PreciseCoder model, frontier models such as GPT five point six Sol and Qwen three point eight Max gain four point six to twelve point four points in edit precision, with no extra training. A linguistic analysis of the thinking traces finds that the reward pushes the model to reason about individual lines, name the buggy ones, and form consistent hypotheses about the bug.

A test that passes says nothing about how a fix got there. Rewarding the path, small and verified edits, gives a debugger that touches less, and a reasoning trace good enough that larger models borrow from it.