Attention Is All You Need

Scope

The 2017 paper by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser and Illia Polosukhin, in its arXiv editions (1706.03762, v1 12 June 2017 through v7 2 August 2023) and its NeurIPS 2017 proceedings edition (Advances in Neural Information Processing Systems 30, paper ID 3058), together with the public NeurIPS reviews; the specific quantitative and architectural claims each edition makes (an architecture without recurrence or convolution, built from attention and feed-forward layers); and the gap between those claims and what the paper is commonly cited as having shown.

This dossier is not a history of the Transformer, a survey of later models, an account of the authors' later careers, or a judgment about which edition is "correct".

Key distinctions

1. "The paper" names several editions and representations. All seven arXiv PDFs were acquired, their embedded v1–v7 labels read, and their SHA-256 digests found to differ. The proceedings PDF and its web abstract are separate representations. A version label or rounded listing size is not a digest; the actual byte comparison is recorded under S02. Different binary hashes do not, by themselves, establish different scientific claims: normalized extracted v6/v7 text differed only in the embedded version/date in this inspection.
2. "The Transformer got 41.8 BLEU on English-to-French" needs a representation and passage. In arXiv v7, the abstract and Table 2 report 41.8, while §6.1 prose reports 41.0. The proceedings HTML abstract reports 41.1; its PDF reports 41.0. Earlier PDFs add further differences: v2 has 41.2 in the abstract, 41.0 in Table 2 and 41.17 in §6.1; v3 has 41.0, 41.0 and 41.17. The complete selected-PDF readings are in S02. These are attributed statements, not an adjudication that every number represents an identical experimental configuration.
3. "28.4 BLEU on English-to-German" is also edition-dependent. The proceedings abstract page reports 27.5; arXiv v7 reports 28.4.
4. arXiv is not the venue. The paper was refereed at NeurIPS 2017; its reviews are public. The arXiv record is the version history, not the peer-reviewed edition.
5. "Introduced attention" is wrong. The paper's background cites earlier attention mechanisms, including Bahdanau et al. (S07). Its architectural claim removes recurrence and convolution from the sequence model while retaining attention and feed-forward layers; the title does not erase those other components.
6. "Proved attention is all you need" is a title, not a result, and "attention only" is not the architecture. The Transformer removes recurrence and convolution from the sequence model; it retains feed-forward layers, residual connections, layer normalization, embeddings and positional encodings alongside attention. The inspected arXiv versions contain two translation tasks and an English constituency-parsing task; the inspected proceedings PDF does not contain that parsing experiment. Its own text leaves restricted attention for very long sequences and other modalities to future work (PC-02 AIAYN-C06, C07).
7. "The paper introduced GPT / BERT / LLMs" is anachronism. Those are later works that used the architecture. The paper claims translation quality, parallelizability and training cost.
8. "Eight Google authors" is approximately right and the footnote is more careful. The paper lists equal contribution with random order and per-author contributions; two authors are marked as having done the work while at Google (one University of Toronto affiliation is listed). Read the footnote in the captured edition rather than paraphrasing it.
9. "Most-cited paper of the century" is a dated, source-specific ranking. Citation counts and rank lists change by database and date; any such sentence needs its source and date (see S08, OPTIONAL).
10. Training cost is version- and configuration-specific. The v7 abstract associates 3.5 days on eight GPUs with the big English-to-French model; the PDF specifies P100 hardware. The v2 abstract instead says 4.5 days. Preserve the edition and experiment rather than describing one training-time number as a property of Transformers (S02).

Questions and evidence limits

• C1 — Edition chronology and representation identity(provisional claim analysis)
• C2 — BLEU values by passage and representation(provisional claim analysis)
• C3 — Individual public reviews(provisional claim analysis)
• C4 — Architectural and experimental scope(provisional claim analysis)
• C5 — Later reputation and unsupported generalization(provisional claim analysis)

Analysis

A citation to the paper does not select one result. In the retained arXiv v7 PDF, the abstract and Table 2 report 41.8 English-to-French BLEU while §6.1 prose reports 41.0. The proceedings HTML abstract and its PDF also differ. The reader can identify which passage supports a precise attribution without deciding that all passages describe identical experimental configurations or choosing an unsupported corrected value.

The inspected representations also delimit the experimental account. All seven retained arXiv PDFs include a constituency-parsing experiment; the inspected proceedings PDF does not. This establishes a difference between the selected documents, not what every reviewer saw or which edition a later citer used. Distinct binary hashes establish distinct selected byte objects, not seven different scientific findings.

The paper’s architectural proposal, earlier attention work, individual reviewer judgments and later engineering work remain separate. Earlier work is cited by the paper; the public reviews contain both favorable assessments and questions. A reviewer’s statement about reimplementation is not an independently verified reproduction record. Later success on other tasks needs the works that report it, rather than being attributed retrospectively to the 2017 experiments.

Questions in the research record

• C-1 — Different English-to-French BLEU statements across and within representations
• C-2 — 28.4 vs 27.5 (English-to-German BLEU)
• C-3 — "Introduced attention" vs "based solely on attention"
• C-4 — Reviewer caveats vs later reputation
• C-5 — Proceedings PDF vs proceedings web abstract
• C-6 — Title vs claim
• C-7 — Training cost vs generality

Why not?

Research account dated 2026-11-06. Descriptions of current or pending states refer to the recorded research date.

Temporal bounds

The selected arXiv editions range from v1 on 12 June 2017 through v7 on 2 August 2023. The proceedings are NeurIPS/NIPS 2017. The seven PDF versions, v1/v7 listings, proceedings PDF/HTML and public review page were retained and read on 6 November 2026; exact passage comparisons were additionally recorded on 7 November. This 11 November candidate makes no new source-retrieval observation.

Version dates and retrieval dates answer different questions. The proceedings HTML reading is a 2026 observation of a proceedings page, not proof of its publication-day 2017 contents. The original unsupported historical-HTML inference remains withdrawn. A dated v5 PDF cannot supply a missing historical HTML snapshot.

The separate PC-02 design dossier remains research_artifact_only and is not superseded or assigned new claim identity here. This publication candidate proposes a distinct Wave 2 dossier target, pending owner choice of route and disposition. Existing Stage AIAYN articles and their architectural HTML bindings remain separate. Later works and optional citation rankings retain their own dates; no current citation ranking or universal scaling conclusion is asserted.