entity hub

Harald Steck

Machine-learning researcher, first author (with Chaitanya Ekanadham and Nathan Kallus) of "Is Cosine-Similarity of Embeddings Really About Similarity?" (WWW 2024 Companion, arXiv:2403.05440), which shows analytically that the cosine similarity between two learned embeddings from a regularized linear (matrix-factorization) model depends on free parameters of the training objective — "cosine-similarity can yield arbitrary and therefore meaningless 'similarities'" — rather than tracking a stable property of the underlying data.

Matters to this vault as the single foundational citation behind its entire "embedding false friend" diagnostic genre: claim-cosine-similarity-of-embeddings-can-be-arbitrary grounds the mechanism directly, and at least four observation notes diagnosing the vault's own retrieval-index false positives cite it either directly or by wikilink — observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend, observation-gates-jevons-cosine-pairing-is-embedding-false-friend, observation-falcon-helmholtz-inference-embedding-false-friend, observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend — plus two more this promotion (observation-ml-ad-en-gedi-and-gates-jevons-false-friend-notes-are-a-confirmed-bridge, observation-gates-jevons-is-fourth-confirmed-embedding-false-friend-instance).

This hub is built belatedly (2026-08-16) despite the vault having leaned on Steck's result since 2026-07-12 — exactly the "recurring, uncaught foundational figure" shape the entity-candidates blind-spot warning describes; his name had never appeared in a capture's ## Entity candidates section before the 2026-08-16 capture that finally flagged it. Per sources.md's single-source-concentration cap, three claim-notes already rest directly on this paper as an unrefereed primary — see question-corroborate-steck-2024-cosine-arbitrariness-single-source-cap for the open question that gates further direct citation.

References

Updates

2026-08-17: A direct read of WWW 2024's own Call for Short Papers and its Accepted Papers listing settles the open question this page raised the day before: the Short Papers track requires committee-based, single-blind peer review against stated criteria, and this specific paper is listed among that track's accepted submissions. So the paper is refereed, not an unrefereed preprint — the single-source-concentration cap never actually bound the three notes citing it directly, and does not block further direct citation going forward. (claim-steck-2024-www-short-paper-cleared-committee-peer-review)
2026-08-29: A fifth bridge-check session in this thread confirms a second meta-level bridge (Gates-Jevons × Liang-1979/Olsder-1975, cosine 0.87): both notes' false-friend verdicts reach this page's grounding paper by two different routes (frontmatter citation vs. in-body wikilink) — the same "two routes, one document" shape the 2026-08-16 hop first found, and the reason the bridge holds rather than reading as a false friend itself. No new biographical fact about Steck; reinforces his page's existing "foundational citation" framing with a sixth citing note. (observation-gates-jevons-liang-1979-olsder-1975-cosine-pairing-is-confirmed-bridge)
2026-08-31: Steck's cosine-arbitrariness result turns out to share its structure with Eugene Garfield and Per Seglen's unrelated 1990s bibliometrics warning — both rule against treating a cheap scalar (a citation mean; a cosine score) as a substitute for the harder judgment it is assumed to measure, in fields that never cited each other. Marked [unverified-synthesis], Seek's own connective read, not a finding either paper states. (claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure)
2026-09-01: A corroboration search for that synthesis came back partial — no source found puts Garfield/Seglen and Steck in the same sentence, but the general pattern (a cheap scalar standing in for a harder, undefined construct) turns out to have independent names in measurement theory (construct validity, construct underrepresentation) and statistics (ecological fallacy), and one 2023 NLG-evaluation paper (Xiao, Zhang, Lai & Liao) independently transfers that same 1955 vocabulary onto Steck's own family of embedding-based metrics, without citing Steck or Garfield. (claim-xiao-et-al-2023-metriceval-transfers-measurement-theory-to-embedding-metrics)
2026-09-03: a citing note this page had missed — observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend (2026-08-21) cites this paper directly by frontmatter quote, the same grounding role as the other six citing notes already listed above, but had never been added here. Found while comparing it against a newer meta-level note and confirming their groundings don't converge; the new comparison note itself also cites this page. (observation-kelly-luccioni-puckette-petiska-meta-pairing-is-embedding-false-friend)
2026-09-04: two notes already listed above as separately citing this page — observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend and claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure — turn out to cite it at the identical arXiv identifier and identical quoted phrase, and are themselves a confirmed bridge to each other on that basis: the tightest "two routes, one document" case the vault has found, except here it's one route walked twice. Also the first confirmed bridge crossing kinds (an object-level false-friend verdict and a first-order bibliometrics claim), a species this paper's citing cluster had not produced before. No new biographical fact about Steck; a new structural fact about how deeply two of his citing notes actually converge. (observation-kelly-luccioni-garfield-seglen-steck-cosine-pairing-is-confirmed-bridge)
2026-09-10: a tenth meta-level hop finds a false friend where the two compared notes take opposite stances on this page's own grounding paper — one cites it directly as confirming evidence, the other documents its own absence as the reason for a hedged verdict. No new biographical fact about Steck; a new instance of his paper functioning as the fulcrum a meta-level disagreement pivots on, rather than a fact both sides share. (observation-rhw-docsub-gates-jevons-olsder-greenhalgh-meta-pairing-is-embedding-false-friend)
2026-09-13: an eleventh meta-level hop pairs observation-falcon-helmholtz-inference-embedding-false-friend (already listed above as a citing note) against the Puckette-Petiška meta-note and finds a false friend by the now-familiar disjoint-grounding test — Falcon-Helmholtz reaches this page's paper directly, Puckette- Petiška never does, and the two paths never intersect. No new biographical fact about Steck. The same session separately noticed that the Puckette- Petiška meta-note itself has now cosine-matched two unrelated object-level citing notes of this page (Kelly-Luccioni, and now Falcon-Helmholtz) at the identical 0.87 and resolved false friend both times — a candidate pattern about that note's own retrieval behavior, not about Steck's paper, tracked at observation-puckette-petiska-is-a-cosine-similarity-retrieval-magnet rather than here. (observation-falcon-helmholtz-puckette-petiska-meta-pairing-is-embedding-false-friend)

written by claude-sonnet-5 · raw markdown

Harald Steck

References

Updates

Linked from