2024 · Conference paper
Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement
Adds two techniques to LLM decompilation of function-level x86-64 GCC assembly into C. Self-Constructed Context Decompilation (sc²dec) recompiles a model’s first compilable output and uses the resulting assembly–source pair as an in-context example for a second attempt, without fine-tuning. Fine-grained Alignment Enhancement (FAE) fine-tunes llm4decompile-6.7b with LoRA on statement-level assembly–C alignments derived from DWARF debug information. Combining both, the authors report 52.41% average re-executability on Decompile-Eval, against 48.60% for the base model.
Scope and limitations
Results are function-level re-executability on Decompile-Eval: 164 HumanEval-derived C problems compiled at O0–O3 and checked by recompilation and HumanEval test cases. Passing these tests is not whole-program or all-input equivalence, and the paper does not evaluate real-world binaries, other architectures or readability. sc²dec helps only when the first output recompiles, so low re-compilability limits its gains; the authors report 1–2% lower results when context is built at a mismatched optimization level, and clang-built contexts scored slightly below GCC. FAE used 10,000 ExeBench functions limited to the standard C library, one training epoch and a single 6.7B base model; the authors did not validate it on larger datasets or models, and it slightly lowered the 1-shot result. GPT-4 and DeepSeek baselines are API snapshots that may not be reproducible.
Why it fits this collection
Adds recompilation feedback and debug-information alignment to end-to-end LLM decompilation, complementing LLM4Decompile and execution-checked decompilation benchmarks.
Publication and review details
- Venue
- Findings of EMNLP 2024
- Status
- Conference paper
- Publication / posting date
- November 2024
- Date source
- The official ACL Anthology citation gives November 2024. Crossref records only the year; the November 12–16 conference dates printed in the PDF are not treated as a publication day.
- First arXiv submission
- June 25, 2024
- arXiv
- 2406.17233
- Reviewed arXiv version
- v2
- DOI
- 10.18653/v1/2024.findings-emnlp.385
The title keeps the published spelling ‘Fined-grained’. The eight-author order follows the Anthology, PDF, Crossref and arXiv. arXiv v2 (2024-10-03) is labeled EMNLP 2024 Findings; v1 was submitted on 2024-06-25. The Anthology and Crossref give pages 6603–6614. The abstract’s ‘approximately 3.90%’ gain differs from the 3.81-point gap between Table 1’s 48.60% base result and 52.41%. At the 2025-02-18 commit, the repository declares MIT, although LICENSE.txt keeps template placeholders for year and holders. It provides an evaluation script and GCC Decompile-Eval data, and links a LoRA adapter for LLM4Binary/llm4decompile-6.7b-v1.5 and a training dataset on Hugging Face, both tagged MIT; the base model and ExeBench keep their own terms. Its README results table adds readability scores and the authors’ later ReF Decompile, and its FAE-only averages differ slightly from the paper (51.07% versus 50.88%). The evaluator uses local gcc and objdump and, with --compiler, sends code to the public Compiler Explorer API.
Sources reviewed
Reviewed against the paper and its primary sources. Cite the original authors using the linked paper or BibTeX.
Primary sources
- https://aclanthology.org/2024.findings-emnlp.385/
- https://aclanthology.org/2024.findings-emnlp.385.pdf
- https://aclanthology.org/2024.findings-emnlp.385.bib
- https://api.crossref.org/works/10.18653/v1/2024.findings-emnlp.385
- https://arxiv.org/abs/2406.17233
- README.md GitHub · AlongWY/sccdec · Revision 22745ea
- LICENSE.txt GitHub · AlongWY/sccdec · Revision 22745ea
- pyproject.toml GitHub · AlongWY/sccdec · Revision 22745ea
- src/sccdec/eval.py GitHub · AlongWY/sccdec · Revision 22745ea
- https://huggingface.co/ylfeng/sccdec-lora
- https://huggingface.co/datasets/ylfeng/sccdec-dataset