{
  "added_at": "2026-10-20",
  "arxiv_id": "2609.17236",
  "arxiv_version_reviewed": "v1",
  "authors": [
    "Muhammad Asjad"
  ],
  "bibtex_url": "https://awesomerea.org/papers/citations/paper-memorization-floor.bib",
  "code_url": null,
  "date_basis": "arXiv v1 submission date; the artifact deposit date is separate.",
  "date_basis_zh": "采用 arXiv v1 首次提交日期；产物存档日期单独区分。",
  "date_precision": "day",
  "description_en": "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.",
  "description_zh": "研究大语言模型改进反编译 C 代码时，标识符恢复来自输入证据还是先验知识。通过同一条目的消融实验和与输出词汇匹配的随机基线，比较教材式函数与新编写函数的命名增益，并进行第二模型复核。",
  "doi": "10.48550/arxiv.2609.17236",
  "entry_url": "https://awesomerea.org/papers/paper-memorization-floor/",
  "entry_url_zh": "https://awesomerea.org/zh/papers/paper-memorization-floor/",
  "evidence_urls": [
    "https://arxiv.org/abs/2609.17236",
    "https://arxiv.org/html/2609.17236v1",
    "https://api.datacite.org/dois/10.48550/arxiv.2609.17236",
    "https://doi.org/10.5281/zenodo.21968878",
    "https://zenodo.org/records/22121473",
    "https://zenodo.org/api/records/21968878"
  ],
  "first_arxiv_date": "2026-09-15",
  "id": "paper-memorization-floor",
  "json_url": "https://awesomerea.org/papers/paper-memorization-floor/record.json",
  "kind": "research",
  "markdown_url": "https://awesomerea.org/papers/paper-memorization-floor/index.md",
  "markdown_url_zh": "https://awesomerea.org/zh/papers/paper-memorization-floor/index.md",
  "metadata_notes": "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.",
  "metadata_notes_zh": "采用单作者的 arXiv v1，未核实正式发表场所。摘要页仍显示待登记，但 DataCite 确认 arXiv DOI 已于 2026-09-16 登记。手稿使用 arXiv 非独占分发许可。论文链接的 Zenodo 概念 DOI 解析到记录 22121473，版本为 1.1.0、日期为 2026-08-27，元数据为 9,639,907 字节的复现 ZIP 声明 CC BY 4.0。",
  "pdf_url": "https://arxiv.org/pdf/2609.17236v1",
  "project_url": "https://zenodo.org/records/22121473",
  "publication_date": "2026-09-15",
  "publication_status": "preprint",
  "record_type": "paper",
  "runtime_tested": false,
  "scope_limitations_en": "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.",
  "scope_limitations_zh": "研究仅含 25 个教材式函数和 20 个作者编写的小函数，使用 x86-64 上的 GCC/Ghidra，每个模型只有一种提示。操作仍可观察，消融也改变惯用模式识别线索，因此有限检出能力下的零结果不能证明记忆或完全不利用输入。没有预登记假设得到确认，正确性指标未通过区分能力门槛；原 T3 逐项产物丢失后重新执行，精度随之改变。多数结果只来自一个改进模型，后续第二模型扩展与最初预登记分开报告。",
  "targets": [
    "binary"
  ],
  "title": "A Memorization Floor for LLM Refinement of Decompiled Code",
  "title_zh": "大语言模型改进反编译代码中的记忆基线",
  "url": "https://arxiv.org/abs/2609.17236",
  "venue": "arXiv",
  "verification": "source-reviewed",
  "verified_at": "2026-10-20",
  "why_fits_en": "Provides controls for interpreting apparently readable or well-named recovered code and cautions against treating naming scores as semantic recovery.",
  "why_fits_zh": "为解释可读性好、命名合理的恢复代码提供对照方法，避免把命名分数当作语义恢复证明。",
  "year": 2026
}