Crossref lists five fake 2025 duplicates of the Transformer paper

What turned up

A 3 September 2026 write-up on dev.to reports that querying Crossref for "Attention Is All You Need" — the 2017 Vaswani et al. paper that introduced the Transformer architecture and, per the post, has drawn more than 140,000 citations — returns five separate records. They share the same title and the same eight authors and are otherwise indistinguishable, yet every one of them lists the publication year as 2025.

All five carry DOIs under the prefix 10.65215 (10.65215/r5bs2d54, 10.65215/ysbyhc05, 10.65215/mdcm8z23, 10.65215/nxvz2v36 and 10.65215/2q58a426). According to the post, that prefix belongs not to a machine learning venue but to an entity calling itself the Shenzhen Medical Academy of Research and Translation, operating through a Chinese preprint host. The author says each DOI was checked directly rather than taken on faith: all five resolved to live, indexed pages, meaning the entries are real Crossref records and not an artifact of search tooling.

Why the authentic record is missing

The reason the duplicates dominate the results, the post explains, is that the genuine 2017 NeurIPS paper has no Crossref DOI in the first place. NeurIPS does not systematically register DOIs for its proceedings, unlike venues covered through the ACL Anthology, so a title search against Crossref alone has nothing legitimate to return. The five fabricated entries are all the query produces.

A wider audit of reference files

The author, who builds a browser-based citation checker called StrictCite, extended the check beyond one paper. Taking ten of the most-read papers on a recent week's Hugging Face Daily Papers listing, they extracted the actual .bib and .bbl reference files from each paper's arXiv source and ran 831 references through thirteen registry checks. More than 300 were flagged, though the post is candid that a meaningful share of this is noise: Semantic Scholar's keyless tier was heavily rate-limited partway through the run, which can make a reference read as unconfirmed for reasons unrelated to the citing paper.

Concrete problems surfaced regardless. One paper's citation year read 1609 — not a date but a fragment of an AAAI DOI (10.1609/...) that a bibliography tool mistook for a year, a pattern the author reports finding independently in three separate papers. In another case, a DOI cited for a diffusion-model paper resolved to an unrelated paper on human-factors modeling for autonomous driving, credited to four authors with no connection to diffusion models.

Official exports from OpenAI and Meta are affected too

The post also examines the BibTeX files that OpenAI and Meta publish for their own papers via arXiv's export endpoint. OpenAI's export for the GPT-4o System Card includes a bare colon listed among the co-authors. Meta's export for "The Llama 3 Herd of Models" splits real people in half: "Guangyi Zhang" becomes "Guangyi" and "Zhang", two nonexistent individuals, with the same fault recurring for "Yu Wang". The likely cause, the post suggests, is the literal string "and" landing inside a name rather than only between names in the generating tooling.

The corruption does not stop at the manuscript. The post states that DataCite, the registry arXiv reports to, has ingested the identical broken author lists as its records of standing. A tool that treats a registry's word as ground truth would mark both entries as correct, because the fault originates at the source rather than downstream of it.

Why it matters

Anything that verifies citations against registry data is exposed to this mechanism. When the correct record is absent — every registry has coverage gaps, and all of them occasionally go down — a single spam duplicate can be the only answer available, and downstream tools will treat it as authoritative. The conclusion the post draws is not that registries are unreliable overall, nor that citation checking is futile, but that a registry confirming a record is not the same thing as that record being correct. Corroboration across independent sources is what carries evidential weight, and a claim resting on exactly one unfamiliar registry deserves skepticism before anyone treats it as settled. Five 2025 impostors sitting where the founding paper of modern language models should be are a small but concrete demonstration of how thin that verification layer can get.