2026 · Preprint
When LLM Decompilers Recompile More and Preserve Less
Decompile-Diverge builds function-level behavioral comparisons from source-derived drivers and fuzzed reference inputs. The study compares reconstructed functions on the same inputs and examines divergences and missing crashes that compilation and supplied tests can miss.
Scope and limitations
The oracle requires reference source, supported inputs and buildable harnesses. It compares a finite corpus and bounded observable state; Matched does not prove all-input equivalence. Nine of 300 GitHub reference functions failed harness builds, leaving 291. Some types/functions are excluded. DeGPT-Qwen is a substituted-backend configuration, not the originally published DeGPT model.
Why it fits this collection
Offers a concrete check on what recompilation and benchmark pass rates do not establish about recovered code.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- September 4, 2026
- Date source
- First arXiv submission date from the reviewed submission history; no formal publication was verified.
- First arXiv submission
- September 4, 2026
- arXiv
- 2609.05370
- Reviewed arXiv version
- v1
Latest reviewed version is arXiv v1, licensed CC BY 4.0. No author-linked code or dataset URL was verified in the reviewed paper and focused artifact search; availability and artifact licenses remain unverified. Reported measurements were not reproduced.
Sources reviewed
Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.