---
license: mit
---

# Precise Debugging Benchmarking (PDB)

⭐ **Neurips'26: Evaluations & Datasets**

📄 [Paper](https://arxiv.org/abs/2604.17338)  · 
💻 [Code](https://github.com/Bill1235813/PDB)  · 
🌐 [Project page](https://precise-debugging-benchmark.github.io/)  · 
🏆 [Leaderboard](https://precise-debugging-benchmark.github.io/leaderboard.html)

**PDB** is an automatic pipeline that turns any coding dataset into a
*debugging* benchmark with fine-grained metrics. Beyond binary unit-test
scores, PDB evaluates a debugger with **edit-level precision** (did the model
touch only the lines it had to?) and **bug-level recall** (did it fix every
fault?). This rewards targeted fixes and penalizes the regeneration behavior
frontier LLMs often fall back on.

> Frontier models like GPT-5.1-Codex and DeepSeek-V3.2-Thinking top unit-test
> leaderboards (>76%) but score at or below 45% on precision: they pass tests
> by rewriting, not repairing. PDB makes that gap measurable.

## Released datasets

| Dataset | Size | Bug granularity | Notes |
|---|---|---|---|
| [PDB-Single](https://huggingface.co/datasets/Precise-Debugging-Benchmarking/PDB-Single) | 5,751 | single line | main evaluation set: tasks not easily solved by 7+ of 9 reference models |
| [PDB-Single-Full](https://huggingface.co/datasets/Precise-Debugging-Benchmarking/PDB-Single-Full) | 7,589 | single line | full initial pool before easy-case filtering |
| [PDB-Wild](https://huggingface.co/datasets/Precise-Debugging-Benchmarking/PDB-Wild) | 484 | multi-line / repository-level | PDB-Multi (256) + SWE-smith repository bugs (228) |
| [PDB-Multi](https://huggingface.co/datasets/Precise-Debugging-Benchmarking/PDB-Multi) | 256 | 2–4 line blocks | BigCodeBench/LiveCodeBench part of PDB-Wild; programs with ≥35 LOC |
| [PDB-Results](https://huggingface.co/datasets/Precise-Debugging-Benchmarking/PDB-Results) | – | – | model outputs and scores |

PDB-Single, PDB-Single-Full and PDB-Multi are derived from [BigCodeBench](https://huggingface.co/datasets/bigcode/bigcodebench)
and [LiveCodeBench](https://huggingface.co/datasets/livecodebench/execution); PDB-Wild adds
[SWE-smith](https://github.com/SWE-bench/SWE-smith) repositories. All are built by the PDB
pipeline and evaluated with precision / recall / unit-test pass rate.

## Citation

```
@article{chai2026pdb,
  title={Precise Debugging Benchmark: Is Your Model Debugging or Regenerating?},
  author={Chai, Miaosen and Zhu, Wang Bill and Wang, Shangshang and Liu, Yejia and Bian, Song and Dong, Honghua and Neiswanger, Willie and Jia, Robin},
  journal={arXiv preprint arXiv:2604.17338},
  year={2026}
}
```

## Contact

Questions / submissions: wangzhu@usc.edu, miaosenc@usc.edu.