PreciseCoder-4B

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

Qwen3.5-4B post-trained with reinforcement learning for precise debugging: given buggy Python code, produce a minimal, correct fix (PreciseCoder project).

Recipe. GDPO with two rewards, unit-test pass and unit-gated F1 (edit-block F1 against the ground-truth diff), weighted 2:1.

Eval:

eval	result
PDB test (1205 BigCodeBench + LiveCodeBench debugging tasks)	unit 0.554 · precision 0.638 · recall 0.728 · F1 0.650 · edit lines 3.0
PDB-Wild (474 tasks: BCB 37 + LCB 209 + SWE-smith 228)	unit 0.276 · precision 0.382 · recall 0.410 · F1 0.377 · edit lines 10.8
GT Debug (1512 clean inputs)	no-op 54.7% · damage 12.1% · unit pass 87.9%

no-op = share of clean (bug-free) inputs returned unchanged; damage = share broken by an edit.

Load with transformers / vLLM (>=0.17) as a standard Qwen3.5 model (thinking enabled by default).

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}
}