{ "added_at": "2026-10-17", "arxiv_id": "2509.22114", "arxiv_version_reviewed": "v1", "authors": [ "Hanzhuo Tan", "Weihao Li", "Xiaolong Tian", "Siyi Wang", "Jiaming Liu", "Jing Li", "Yuqun Zhang" ], "bibtex_url": "https://awesomerea.org/papers/citations/paper-sk2decompile.bib", "code_url": "https://github.com/albertan017/LLM4Decompile/tree/85b364bf093eb2ee4f3687cfe38a203fca89f23e/sk2decompile", "date_basis": "Publication year stated by the official ICLR proceedings; no finer publication date is asserted.", "date_basis_zh": "采用 ICLR 官方论文集标明的出版年份，不推定更精确的出版日期。", "date_precision": "year", "description_en": "Splits LLM decompilation into two phases: a Structure Recovery model turns decompiler pseudocode into an identifier-free IR that keeps control flow and data structures, trained with compiler-based RL rewards, and an Identifier Naming model then restores names under a semantic-similarity reward. The authors report a 21.6% average re-executability gain over GPT-5-mini on HumanEval and a 29.4% average R2I improvement over Idioms on GitHub2025.", "description_zh": "将 LLM 反编译分为两个阶段：结构恢复模型把反编译器伪代码转换为保留控制流和数据结构、但不含标识符的中间表示，并以编译器反馈作为强化学习奖励；标识符命名模型再以语义相似度为奖励恢复名称。作者报告，其在 HumanEval 上的平均可重新执行率比 GPT-5-mini 高 21.6%，在 GitHub2025 上的平均 R2I 比 Idioms 提升 29.4%。", "doi": null, "entry_url": "https://awesomerea.org/papers/paper-sk2decompile/", "entry_url_zh": "https://awesomerea.org/zh/papers/paper-sk2decompile/", "evidence_urls": [ "https://proceedings.iclr.cc/paper_files/paper/2026/hash/9132583dd869acb0cafeb0ed25eb031a-Abstract-Conference.html", "https://proceedings.iclr.cc/paper_files/paper/2026/file/9132583dd869acb0cafeb0ed25eb031a-Paper-Conference.pdf", "https://openreview.net/forum?id=jSQPqdoidy", "https://iclr.cc/virtual/2026/poster/10007876", "https://arxiv.org/abs/2509.22114", "https://github.com/albertan017/LLM4Decompile/blob/85b364bf093eb2ee4f3687cfe38a203fca89f23e/sk2decompile/README.md", "https://github.com/albertan017/LLM4Decompile/blob/85b364bf093eb2ee4f3687cfe38a203fca89f23e/LICENSE", "https://github.com/albertan017/LLM4Decompile/blob/85b364bf093eb2ee4f3687cfe38a203fca89f23e/LICENSE-MODEL", "https://huggingface.co/LLM4Binary/sk2decompile-struct-6.7b/blob/3c173bebc9cf1d3a06b0607a743e0f0084d2a50b/README.md", "https://huggingface.co/LLM4Binary/sk2decompile-ident-6.7b/blob/62119d483580bff1a454ec2c774bad7c6086969e/README.md" ], "first_arxiv_date": "2025-09-26", "id": "paper-sk2decompile", "json_url": "https://awesomerea.org/papers/paper-sk2decompile/record.json", "kind": "research", "markdown_url": "https://awesomerea.org/papers/paper-sk2decompile/index.md", "markdown_url_zh": "https://awesomerea.org/zh/papers/paper-sk2decompile/index.md", "metadata_notes": "Authors follow the proceedings PDF; the proceedings landing page writes the first author's given name as 'hanzhuo'. OpenReview lists the paper as an ICLR 2026 Poster. The September 2025 arXiv v1 is separate from ICLR 2026 publication, and the 32-page proceedings version adds appendix experiments. Code is the sk2decompile directory of the LLM4Decompile repository. License signals differ by source: the repository LICENSE is MIT, both 6.7B model cards are tagged MIT, and the repository also carries a DeepSeek LICENSE-MODEL; both models are initialized from LLM4Decompile-6.7B. No single license is asserted.", "metadata_notes_zh": "作者以论文集 PDF 为准；论文集网页将第一作者的名字写作小写的“hanzhuo”。OpenReview 将其列为 ICLR 2026 Poster。2025 年 9 月的 arXiv v1 与 ICLR 2026 发表分开记录，32 页的论文集版本增加了附录实验。代码位于 LLM4Decompile 仓库的 sk2decompile 目录。各来源的许可信息不同：仓库 LICENSE 为 MIT，两个 6.7B 模型卡均标注 MIT，仓库另附 DeepSeek LICENSE-MODEL；两个模型均从 LLM4Decompile-6.7B 初始化。本条不认定统一许可。", "pdf_url": "https://proceedings.iclr.cc/paper_files/paper/2026/file/9132583dd869acb0cafeb0ed25eb031a-Paper-Conference.pdf", "project_url": null, "publication_date": null, "publication_status": "conference-paper", "record_type": "paper", "runtime_tested": false, "scope_limitations_en": "Training and main evaluation use function-level C compiled for x86-64 Linux with GCC and Clang at -O0 to -O3, with IDA Pro pseudocode as model input. Re-executability is measured only on HumanEval and MBPP; ExeBench and GitHub2025 use R2I readability and a GPT-5-mini judge. Appendix experiments extend to BringUpBench projects, macOS arm64 and Windows MSVC builds, and Obfuscator-LLVM, but the models are trained on C only and the paper expects them to degrade on features without C equivalents, such as C++ templates.", "scope_limitations_zh": "训练和主要评测均使用函数级 C 代码，以 GCC 和 Clang 在 -O0 至 -O3 下编译为 x86-64 Linux 二进制，并以 IDA Pro 伪代码作为模型输入。可重新执行率仅在 HumanEval 和 MBPP 上测量；ExeBench 与 GitHub2025 采用 R2I 可读性指标和 GPT-5-mini 评审。附录实验扩展到 BringUpBench 项目、macOS arm64 与 Windows MSVC 构建以及 Obfuscator-LLVM，但模型仅以 C 代码训练，论文预计遇到 C++ 模板等没有 C 对应物的语言特性时效果会下降。", "targets": [ "binary" ], "title": "SK2Decompile: LLM-based Two-Phase Binary Decompilation from Skeleton to Skin", "title_zh": "SK2Decompile：从骨架到皮肤的大语言模型两阶段二进制反编译", "url": "https://proceedings.iclr.cc/paper_files/paper/2026/hash/9132583dd869acb0cafeb0ed25eb031a-Abstract-Conference.html", "venue": "ICLR 2026", "verification": "source-reviewed", "verified_at": "2026-10-17", "why_fits_en": "Adds a recent LLM decompiler that separates structural correctness from identifier recovery, extending the LLM4Decompile line of work.", "why_fits_zh": "补充一种较新的 LLM 反编译方法，将结构正确性与标识符恢复分开处理，延续了 LLM4Decompile 系列工作。", "year": 2026 }