# Decompile-Bench: Million-Scale Binary-Source Function Pairs for Real-World Binary Decompilation

> Introduces two million C/C++ binary-source function pairs using a Compile-Trace-Filter pipeline, with an evaluation suite spanning translated HumanEval/MBPP tasks and recent GitHub functions.

- Authors: Hanzhuo Tan; Xiaolong Tian; Hanrui Qi; Jiaming Liu; Zuchen Gao; Siyi Wang; Qi Luo; Jing Li; Yuqun Zhang
- Venue: NeurIPS 2025 Datasets and Benchmarks
- Status: Conference paper
- Publication / posting date: 2025 (exact date not verified)
- Date source: Publication year in the official NeurIPS proceedings and publisher-deposited Crossref metadata; no finer publication date is asserted.
- First arXiv submission: May 19, 2025
- arXiv: 2505.12668
- Reviewed arXiv version: v2
- DOI: 10.52202/085713-0177
- Sources reviewed: 2026-10-16
- Added to the catalog: 2026-10-17
- Original paper: <https://papers.nips.cc/paper_files/paper/2025/hash/079cf13ae174c31f148207d94d213bdc-Abstract-Datasets_and_Benchmarks_Track.html>
- PDF: <https://papers.nips.cc/paper_files/paper/2025/file/079cf13ae174c31f148207d94d213bdc-Paper-Datasets_and_Benchmarks_Track.pdf>
- Code: <https://github.com/albertan017/LLM4Decompile>
- BibTeX: <https://awesomerea.org/papers/citations/paper-decompile-bench.bib>
- Catalog page: <https://awesomerea.org/papers/paper-decompile-bench/>
- Record JSON: <https://awesomerea.org/papers/paper-decompile-bench/record.json>
- Chinese version: [简体中文](https://awesomerea.org/zh/papers/paper-decompile-bench/index.md)

## Scope and limitations

This is function-level decompilation. Re-execution tests cover HumanEval/MBPP; GitHub2025 provides no unit tests. Inlining can prevent one-to-one source correspondence, and the fine-tuning study uses 10% of the corpus. These results do not establish whole-program recovery.

## Why it fits this collection

Adds a training corpus and evaluation artifact for LLM decompilation, distinct from Gao et al.’s existing DecompileBench entry.

## Publication and review notes

Author order follows the final PDF, which puts Zuchen Gao before Siyi Wang; the proceedings landing page and Crossref reverse them. arXiv v2 was revised on 2025-10-19. The artifact is the decompile-bench subdirectory of the existing LLM4Decompile repository. Code has an MIT license; the dataset declares CC0-1.0, while model weights have separate terms. These declarations do not replace source-project license review.

## Primary sources

1. <https://papers.nips.cc/paper_files/paper/2025/hash/079cf13ae174c31f148207d94d213bdc-Abstract-Datasets_and_Benchmarks_Track.html>
2. <https://papers.nips.cc/paper_files/paper/2025/file/079cf13ae174c31f148207d94d213bdc-Paper-Datasets_and_Benchmarks_Track.pdf>
3. <https://doi.org/10.52202/085713-0177>
4. <https://api.crossref.org/works/10.52202/085713-0177>
5. <https://arxiv.org/abs/2505.12668>
6. <https://arxiv.org/html/2505.12668v2>
7. <https://github.com/albertan017/LLM4Decompile/blob/85b364bf093eb2ee4f3687cfe38a203fca89f23e/decompile-bench/readme.md>
8. <https://github.com/albertan017/LLM4Decompile/blob/85b364bf093eb2ee4f3687cfe38a203fca89f23e/LICENSE>
9. <https://github.com/albertan017/LLM4Decompile/blob/85b364bf093eb2ee4f3687cfe38a203fca89f23e/LICENSE-MODEL>
10. <https://huggingface.co/datasets/LLM4Binary/decompile-bench/blob/4b708c2211cd7d4af403675db56322aa4ed7050c/README.md>
11. <https://huggingface.co/datasets/LLM4Binary/decompile-bench/blob/4b708c2211cd7d4af403675db56322aa4ed7050c/LICENSE>

## About this entry

Reviewed against the paper and its primary sources. Cite the original authors using the linked paper or BibTeX.
Original editorial content: MIT. Linked projects, papers and assets retain their own rights and licenses.