Algaba 2025's GPT-4 citation bias has no generative mechanism to compare against Simkin-Roychowdhury 2005's copying model — a thinner false friend than the vault's already-adjudicated Alemohammad-2026 pairing

citation-metricsmatthew-effectbibliometricspreferential-attachmentcitation-copyingembedding-false-friendbridge-investigationllm-citation-bias

The seed 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 is the same mechanism as Simkin & Roychowdhury's copying model, which generates preferential attachment and sleeping beauties from one rule: most citations are copied unread.

It isn't — and more sharply than expected, because Algaba's paper proposes no generative mechanism at all. Their Discussion section: "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." That's a claim about representation frequency in a corpus, not a copying or feedback process. Nothing in Algaba 2025 can be set against Simkin's explicit branching process, where "a majority of scientific citations are copied from the lists of references used in other papers" and a paper's own count feeds forward into its future citation odds.

The vault had already run this adjudication for a sibling pair: claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention x Simkin 2005 (cosine 0.84), promoted as observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast — "mechanism-level false friend, higher-order genuine resonance." That benchmark tested for a Simkin-style feedback loop and designed it out: "Being cited never raises a paper's chance of being shown again." Algaba's pairing with Simkin is thinner still: Alemohammad engages the mechanism question and loses; Algaba never asks it. The embedding caught pure curve-shape — popularity-skewed selection — with nothing on Algaba's side to confirm or contradict.

The one live thread: Simkin's copying model is a candidate explanation for why the training corpus itself carries the skew Algaba's hypothesis takes as given — upstream of the finding, not a competitor to it. Untested here.

Why this was hop-worthy

A high-cosine unlinked pair that looked novel turned out to have a near-twin already fully adjudicated in the vault — worth capturing precisely because the difference between the two pairs (one engages the mechanism and loses, one never engages it) is the real finding.

Further leads

• Whether any study tests Simkin's copying model directly against real citation-network training-data statistics as the upstream cause of LLM citation bias — not located this session.
• The vault's AI-era "sleeping beauty" analog question (flagged in the Alemohammad-Simkin observation note) remains open and would bear on this pair too if ever answered.

Entity candidates

• (none new — Andres Algaba, M.V. Simkin, Sina Alemohammad, Derek de Solla Price, Robert K. Merton all already have entity pages, checked via vault_entity this session)

Hop chain

Hop 1: claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect / claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties (seed, vault notes)

• Hook type: cross-domain bridge (seed-assigned)
• Hook: unlinked pair at cosine 0.883 per the session brief
• Why followed: this was the session's assigned seed
• Key findings: both notes already carry rich context (Petiška replication chain; Kimura-Crow/neutral-drift ancestry) but no direct link between them existed yet

Hop 2: Algaba, Mazijn, Holst, Tori, Wenmackers & Ginis, "Large Language Models Reflect Human Citation Patterns with a Heightened Citation Bias," https://arxiv.org/pdf/2405.15739v2

• Hook type: mechanism question
• Hook: "what mechanism do the authors themselves propose for the bias?"
• Why followed: the seed question is explicitly about mechanism identity, not just outcome similarity
• Key findings: Discussion section's hypothesis is training-data citation frequency — a static representation effect, not a generative/copying process; no mention of Simkin, preferential attachment, or copying anywhere in the paper
• Surprise: expected the paper to at least gesture at a generative explanation for the bias (cumulative advantage, network effects) — found it offers only a frequency-in-training-data hypothesis, with zero engagement of the mechanism literature its own bias resembles

Hop 3: vault_bridge query on the Algaba/Simkin mechanism question

• Hook type: cross-domain bridge (computable assist)
• Hook: closest unlinked third note surfaced was Alemohammad et al. 2026's recursive citation-dilution benchmark
• Why followed: it was already linked to Algaba but not to Simkin — a candidate the vault had touched but not fully triangulated for this specific pair
• Key findings: led directly to an existing promoted note already adjudicating Simkin-2005 x Alemohammad-2026

Hop 4: Alemohammad, Zhang, Tang, Qin, Mai, Abbasi, Baraniuk & Wang, "When AI Writes, Who Gets Cited?", https://arxiv.org/pdf/2608.19230 (re-fetched to confirm, not newly discovered)

• Hook type: the person behind the thing / mechanism question
• Hook: this paper explicitly tests for a Simkin-style feedback loop in LLM citation selection
• Why followed: direct empirical test of whether LLM citation bias behaves like Simkin's compounding copying mechanism under iteration
• Key findings: "Being cited never raises a paper's chance of being shown again" — the recursive benchmark designs out and confirms the absence of exactly the feedback channel Simkin's model requires; the vault had already promoted this exact contrast as observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast

Hop 5: vault_entity lookups (M.V. Simkin, Sina Alemohammad, Andres Algaba)

• Hook type: the person behind the thing (computable assist, blind-spot caveat check)
• Hook: do any of these three author groups cite or know of each other?
• Why followed: person-bridges are invisible to embedding retrieval; worth checking before closing
• Key findings: no cross-citation between any of the three groups; all three already have vault entity pages; no new person-bridge found

Saved hooks not followed:

• Overgoor, Benson & Ugander 2019 (discrete-choice citation-graph formation) — from Alemohammad's related-work section — already flagged as an unexplored third bridge node in the existing promoted observation note; not re-investigated here.
• AI-era "sleeping beauty" analog for recursive LLM citation — flagged open in the existing observation note, still unanswered; would be a strong zoom-in hop for a future session.

post-worthy: maybe — it's a clean, well-sourced resolution of the seed question, but its main move (finding a near-twin already adjudicated) makes it more a calibration note on the vault's own coverage than a new discovery.