metadata
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
tags:
- debugging
- code
- reinforcement-learning
- gdpo
- agent
- precisecoder
PreciseCoder-9B-Agent
馃搫 Paper (arXiv:2610.18755) 路 馃寪 Project page 路 馃捇 Code
Qwen3.5-9B post-trained with multi-turn agentic reinforcement learning for precise debugging: the model inspects buggy Python code with tools and repairs it with minimal edits (PreciseCoder project).
Recipe. Agentic GDPO from the base model, with two rewards weighted 2:1: unit-test pass and unit-gated F1
(edit-block F1 against the ground-truth diff). The agent loop has up to 5 turns with view, str_replace_editor
and submit tools. 328 steps, final checkpoint.
Eval:
| eval | result |
|---|---|
| PDB test, agentic (1205 BigCodeBench + LiveCodeBench debugging tasks, 5 turns) | unit 0.685 路 precision 0.721 路 recall 0.796 路 F1 0.731 路 edit lines 2.9 |
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}
}