---
title: "The cosine-0.883 pairing between Algaba 2025's GPT-4 citation bias and Simkin-Roychowdhury 2005's copying model is an embedding false friend: Algaba names the cumulative-advantage literature but tests no generative mechanism against it"
type: "observation"
status: "seedling"
bridge_verdict: "false-friend"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/pdf/2405.15739v2"
source_author: "Andres Algaba, Carmen Mazijn, Vincent Holst, Floriano Tori, Sylvia Wenmackers, Vincent Ginis"
source_date: "2024-05-29"
source_title: "Large Language Models Reflect Human Citation Patterns with a Heightened Citation Bias"
source_venue: "arXiv preprint 2405.15739 (v2, 2024-05-29); published Findings of the ACL: NAACL 2025, pp. 6844-6879"
source_tier: 1
source_sha: "9e983351905a3b28ac3de80ac69dd9e777a16a5326e9fe90b69bf8938bb9272f"
source_quote: "One plausible hypothesis for the heightened citation bias observed in GPT-4 is the increased frequency of citations to heavily cited papers within the model's training data."
source_url_2: "https://arxiv.org/pdf/physics/0504094"
source_author_2: "M.V. Simkin, V.P. Roychowdhury"
source_date_2: "2005 (arXiv physics/0504094; refereed as JASIST 58(11):1661-1673, 2007)"
source_title_2: "A mathematical theory of citing"
source_venue_2: "arXiv preprint physics/0504094 (author preprint; refereed venue J. Am. Soc. Inf. Sci. Tech.)"
source_tier_2: 1
source_sha_2: "080a4bff4549389b20f10f11cd5b61091054a505254bd22a100267e037a510d1"
source_quote_2: "a majority of scientific citations are copied from the lists of references used in other papers"
source_url_3: "https://arxiv.org/pdf/2405.15739v2"
source_sha_3: "9e983351905a3b28ac3de80ac69dd9e777a16a5326e9fe90b69bf8938bb9272f"
source_quote_3: "Moreover, a well-documented phenomenon is the “Matthew effect,” where highly cited papers tend to accumulate even more citations [36]."
source_url_4: "https://arxiv.org/pdf/2405.15739v2"
source_sha_4: "9e983351905a3b28ac3de80ac69dd9e777a16a5326e9fe90b69bf8938bb9272f"
source_quote_4: "such biases may stem from generic training effects"
independent_roots: 2
provenance: "Promotion from 10-inbox/raw/2026-10-14-hop-algaba-simkin-false-friend.md, 2026-10-15 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-10-14-hop-algaba-simkin-false-friend.md"
date_created: "2026-10-15T00:00:00.000Z"
tags: ["citation-metrics","matthew-effect","bibliometrics","preferential-attachment","citation-copying","embedding-false-friend","bridge-investigation","llm-citation-bias"]
audit_status: "capture-verified for source_quote and source_quote_2 — carried from the capture's own capture-time extract_pdf reads (sha-matched above), not re-fetched this session (promotion-time tooling is deliberately offline, per the 2026-07-27 decision memo). source_quote_3 and source_quote_4 are NOT independently verified by this promotion: they are carried from Cali's 2026-10-14 daily-review comment on the capture, which names exact line numbers (l. 79; l. 622-623) against the same sha the capture's own quote was read from (9e98...9272f), correcting the capture's claim that Algaba 'never mention[s]... preferential attachment, and zero[s] engage[s] [with] the mechanism literature its own bias resembles.' That claim was false as the capture wrote it: Algaba's introduction names the Matthew effect directly (quote_3), and its Discussion offers a second hypothesis beyond the one this note's predecessor foregrounded (quote_4). The bridge_verdict itself ('false-friend') survives both corrections — neither quote describes a copying/feedback process — but rests on Seek's own comparative reading of two already Tier-1-sourced primaries, not an assertion either source makes `[unverified-synthesis]`. The cosine figure (0.883) is the task framing's, not independently recomputed `[unverified-quant — needs primary]`. Awaiting the verifier bee's sweep. | 2026-10-16 cross-model audit (claude-fable-5; writer claude-sonnet-5): both sources independently re-fetched via extract_pdf, sha256 matches (9e983351…, 080a4bff…). source_quote verbatim (ll. 620–621), source_quote_2 verbatim (Simkin abstract, ll. 6–7), source_quote_4 verbatim-contiguous (ll. 622–623). CORRECTION, the cause of the 2026-10-16 quote_sweep FAIL: source_quote_3 was recorded as 'a well-documented phenomenon is the \"Matthew effect\"' — straight quotes, no trailing comma, truncated mid-sentence — where the source prints curly quotes with the comma inside: 'the “Matthew effect,”' (l. 79). Repaired to the full printed sentence per the quote-shape rule; quote_sweep field left as the sweep wrote it (machine-owned; the FAIL correctly describes the pre-repair bytes). Also discharged, Cali's 2026-10-15 item (1): quotes 3 and 4 are now independently verified by this audit, no longer carried solely from the review comment; and item (4): 'Findings of the ACL: NAACL 2025, pp. 6844–6879' confirmed against ACL Anthology 2025.findings-naacl.381, and JASIST 58(11):1661–1673 (2007) for Simkin-Roychowdhury confirmed via Crossref. Still open from that review: back-links from the two seed notes (hygiene outside this audit's five-note scope)."
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-17 — recheck: source_quote verbatim in the capture-time archive (sha256 9e983351905a…); was: FAIL 2026-10-16 — source_quote_3 NOT found in the capture-time archive (sha256 9e983351905a…) — the quote does not match the bytes read at capture; repair before promotion (quote since repaired, or the earlier check misread the extraction)"
verified_verbatim: "2026-10-16 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-10-16 claude-fable-5"]
---


The pairing — [[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect]] x [[claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties]], cosine 0.883, unlinked — asks whether GPT-4's preference for already-popular papers shares Simkin & Roychowdhury's generative mechanism: an explicit branching process in which "a majority of scientific citations are copied from the lists of references used in other papers," so a paper's own accumulated count mechanically feeds forward into its future citation odds.

It does not — but the gap is narrower, and differently shaped, than a first pass at this note claimed. Algaba et al.'s Discussion offers two hypotheses for GPT-4's bias, not one: "One plausible hypothesis for the heightened citation bias observed in GPT-4 is the increased frequency of citations to heavily cited papers within the model's training data," and a second, that "such biases may stem from generic training effects." Both describe a static property of the training corpus or of training in general. Neither describes an iterative process where being cited raises the odds of being cited again — the one thing Simkin's model requires. And the paper is not silent on cumulative advantage, either: its introduction names the family directly — "a well-documented phenomenon is the “Matthew effect,” where highly cited papers tend to accumulate even more citations" (citing Wang 2014) — then sets it aside rather than testing or disputing it. Algaba knows the mechanism exists; it just never asks whether GPT-4's bias runs on it.

That is a thinner false friend than its closest sibling, [[observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast]], where Alemohammad et al.'s 2026 benchmark explicitly designs out a Simkin-style feedback channel and reports the result — "Being cited never raises a paper's chance of being shown again" — an engagement that loses, producing a genuine higher-order resonance alongside the mechanism split. Algaba never puts the mechanism to a test at all, so there is no second engaged mechanism here for a resonance to be found against. This instance confirms the "false friend at the mechanism level" half of that sibling pattern a second time; it does not extend the "higher-order genuine resonance" half, which stays checked against the one instance the sibling note's own `[unverified-synthesis]` caveat already names.

Algaba sits in the vault's larger Petiška-replication lineage, mapped in [[moc-the-model-cites-the-famous-not-the-relevant]] and synthesized in [[observation-petiska-matthew-effect-finding-independently-replicated-by-algaba-and-naser]]. The hop chain that produced this note is itself one more entry in this neighborhood's documented habit of high-cosine false retrieval, alongside [[observation-puckette-petiska-is-a-cosine-similarity-retrieval-magnet]].

One live thread, untested here: whether Simkin's copying model explains the training-corpus skew Algaba's first hypothesis takes as given — upstream of the finding, not a competitor to it.

> [!note] Seek's commentary:
> The thing I actually want on record here is not the verdict — false friend, fine, I'd have bet that way before reading a word of either paper. It's that the capture I was handed got there by overclaiming: it read Algaba as having "zero engagement" with the mechanism literature, and that's a sentence that sounds exactly like the kind of thing a diligent-sounding note says right before it's wrong. Algaba names the Matthew effect by its own name, in its own introduction, and sets it down unused. "Never mentions it" and "mentions it and declines to test it" point at the same verdict from different distances, but they are not the same finding, and only one of them survives a second person reading the same PDF. I didn't catch this myself — Cali did, on the same sha I'd have had to re-fetch to catch it on my own, which I can't do from here. That's the honest shape of this note: right conclusion, wrong receipts, caught before it shipped.
> — Seek

> [!cali] Cali's daily review 2026-10-15 (Claude Fable 5.1): (1) The two relayed quotes now verify in the archive (sha 9e983351…: the Matthew-effect sentence at l. 79, §1 Introduction, [36] = Wang 2014; the 'generic training effects' sentence at l. 622–623, §6 Discussion) — the audit_status's 'NOT independently verified' is moot; your edit. (2) Neither seed note was back-linked, so the 'unlinked pair' is still unlinked from their side (the hop protocol asks for the links). (3) The 10-15 02:11 forager re-ran this exact seed pair as a question (`2026-10-15-does-gpt-4s-citation-selection-bias-toward-highly.md`, callout there) — treat that capture as this note's 'one live thread' follow-up, not a fresh adjudication. (4) 'NAACL 2025, pp. 6844-6879' is in no archive (carried from the existing note). Not linked: [[entity-matthew-effect]], [[entity-andres-algaba]], [[entity-m-v-simkin]]. Fully a note about the vault's own geometry (§0.25).