PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards

📄 Paper (arXiv:2610.18755) · 🌐 Project page · 💻 Code

Unit-test rewards are the standard way to post-train code models, but a unit test only checks whether the final program passes. Training on them raises pass rates while teaching models to edit more than needed and to hallucinate bugs in code that is already correct. PreciseCoder adds a reward for semantically verified, localized edits. On 4B, 9B and 27B models, it cuts the average edit from 7–15 lines to 3 at a pass rate on par with unit-test-only training. The precision carries over to repository-level repair and, through PreciseCoder's reasoning traces, to frontier models.

Models

• PreciseCoder-4B, from Qwen3.5-4B
• PreciseCoder-9B, from Qwen3.5-9B
• PreciseCoder-27B, from Qwen3.6-27B

Data

• pdb-test: the held-out test split of PDB, the Precise Debugging Benchmark

Citation

@article{precisecoder,
  title   = {{PreciseCoder}: Faithful and Generalizable Debugging Beyond Unit-Test Rewards},
  author  = {Zhan, Yunkai and Chai, Miaosen and Wang, Shangshang and Zhong, Jike and Yang, Yuqing and Fu, Deqing and Chang, Serina and Zhu, Wang Bill},
  journal = {arXiv preprint arXiv:2610.18755},
  year    = {2026}
}