2026 · Preprint
A Memorization Floor for LLM Refinement of Decompiled Code
Studies whether LLM refinement of decompiled C recovers identifiers from input evidence or prior knowledge. Within-item ablations and output-matched chance baselines test naming gains on textbook and newly authored functions, including a second-model replication.
Scope and limitations
The study uses 25 textbook and 20 author-written small functions, GCC/Ghidra on x86-64, and one prompt per model. Operations remain visible and the ablation also changes idiom-recognition cues, so its bounded null does not prove memorization or absence of input use. No registered hypothesis was confirmed; the correctness measure failed its discrimination gate; original T3 per-item artifacts were lost and re-executed, changing precision. Most findings use one refiner, and the later second-model extension is distinct from the initial preregistration.
Why it fits this collection
Provides controls for interpreting apparently readable or well-named recovered code and cautions against treating naming scores as semantic recovery.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- September 15, 2026
- Date source
- arXiv v1 submission date; the artifact deposit date is separate.
- First arXiv submission
- September 15, 2026
- arXiv
- 2609.17236
- Reviewed arXiv version
- v1
- DOI
- 10.48550/arxiv.2609.17236
Single-author arXiv v1; no formal venue was verified. DataCite confirms the arXiv DOI was registered on 2026-09-16 despite the abstract page’s pending label. The manuscript uses arXiv’s non-exclusive distribution license. The paper’s Zenodo concept DOI resolves to record 22121473, version 1.1.0, dated 2026-08-27, declaring CC BY 4.0 for a 9,639,907-byte replication ZIP. Only deposit metadata was inspected, not the archive or its per-file licenses. No experiments were run.
Sources reviewed
Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.