2026 · Preprint
Binary Decompilation LLM with Feedback-Driven Multi-Turn Refinement
AutoDecompiler trains function-level decompilation models to revise assembly- or pseudocode-derived outputs using compilation, execution and I/O feedback. Its reinforcement-learning design combines diagnostic feedback with rewards for valid code, tested behavior, syntax and data-flow structure.
Scope and limitations
Evaluation uses C adaptations of HumanEval and MBPP with validation feedback; finite tests do not establish general equivalence. HumanEval also informs reward-weight selection, and base-model pretraining exposure cannot be ruled out. Refinement is context-limited, whole-program dependencies are not recovered, and unseen toolchains or protected binaries remain limitations.
Why it fits this collection
Adds reinforcement-learning-trained use of program feedback, with separately recorded first submission and latest revision.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- June 15, 2026
- Date source
- First arXiv submission on 2026-06-15; the reviewed v2 revision is 2026-10-03, not a separate publication.
- First arXiv submission
- June 15, 2026
- arXiv
- 2606.16162
- Reviewed arXiv version
- v2
First submitted 2026-06-15; reviewed v2 is dated 2026-10-03. The arXiv record uses the shorter catalog title, while the v2 PDF/HTML use a longer title beginning “AutoDecompiler: Reinforcement-Learning-Trained”; eight authors match. Zenodo version 23120286 lists code/data archives under CC BY 4.0; only metadata and its small README were read. The inspected Hugging Face model has no license metadata. No weights or archives were downloaded or run.
Sources reviewed
Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.