PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards
Unit-test rewards made code models pass more tests, but also pushed them toward unnecessary edits.
The paper introduces PreciseCoder, which adds a semantically verified signal for localized debugging changes. In the reported runs, it cut average edit lines from 7-15 to 3 while keeping pass rates comparable to unit-test-only training. The authors say the method generalized from single-line bugs to repository-level repair, producing smaller patches and resolving more issues. They also report that traces from a 9B PreciseCoder model improved edit precision in larger frontier models without additional training. ArXiv · AI/CL/LG's note
The paper introduces PreciseCoder, which adds a semantically verified signal for localized debugging changes. In the reported runs, it cut average edit lines from 7-15 to 3 while keeping pass rates comparable to unit-test-only training. The authors say the method generalized from single-line bugs to repository-level repair, producing smaller patches and resolving more issues. They also report that traces from a 9B PreciseCoder model improved edit precision in larger frontier models without additional training. ArXiv · AI/CL/LG's note
score 6