2026 · Conference paper

SK2Decompile: LLM-based Two-Phase Binary Decompilation from Skeleton to Skin

Hanzhuo Tan; Weihao Li; Xiaolong Tian; Siyi Wang; Jiaming Liu; Jing Li; Yuqun Zhang

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.

Scope and limitations

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.

Why it fits this collection

Adds a recent LLM decompiler that separates structural correctness from identifier recovery, extending the LLM4Decompile line of work.

Publication and review details

Venue: ICLR 2026
Status: Conference paper
Publication / posting date: 2026 (exact date not verified)
Date source: Publication year stated by the official ICLR proceedings; no finer publication date is asserted.
First arXiv submission: September 26, 2025
arXiv: 2509.22114
Reviewed arXiv version: v1

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.

Sources reviewed 2026-10-17

Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.

Scope and limitations

Why it fits this collection

Publication and review details

Primary sources