---
id: "20260928-0204-bipartite-two-things-this"
title: "Bipartite hop: does Alemohammad 2026's recursive dilution-concentration finding share a mechanism with Simkin & Roychowdhury 2005's random-copying model, or is the cosine-0.84 resemblance a false friend?"
type: "capture"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-09-28T00:00:00.000Z"
provenance: "Batch research run, 2026-09-28, investigating the unlinked pair claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention x claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties (cosine 0.84)"
derived_from: []
tags: ["citation-analysis","preferential-attachment","citation-monoculture","model-collapse","bibliometrics","mechanism-comparison"]
source_url: "https://arxiv.org/pdf/2608.19230"
source_title: "When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models"
source_author: "Sina Alemohammad, Denghui Zhang, Bolong Tang, Anthony Qin, Gengchen Mai, Ahmed Abbasi, Richard Baraniuk, Zhangyang Wang"
source_date: "2026-08-03"
source_venue: "arXiv preprint 2608.19230 [cs.DL]"
source_tier: 1
source_sha: "fa6cfef20674c458d602e81f7ff17913371cc3f6adc5af141a9a306668d28689"
source_url_2: "https://arxiv.org/pdf/physics/0504094"
source_title_2: "A mathematical theory of citing"
source_author_2: "M.V. Simkin, V.P. Roychowdhury"
source_date_2: "2005 (arXiv physics/0504094; published refereed as JASIST 58(11):1661-1673, 2007)"
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"
recovered: true
recovered_on: "2026-10-01T00:00:00.000Z"
recovered_from: "Seek session transcript 38f036df-2e4e-4e2b-a38d-5d3cd9447cbf.jsonl on the mini - the session's Write at 2026-09-28 09:11 UTC, replayed verbatim"
recovery_note: "Overwritten the same night: another bipartite session was told to write the same filename (2026-09-28-bipartite-two-things-this-vault-knows-in-different.md). Collision fixed in seek-code c516ae8. Restored unpromoted at Cali's instruction, 2026-10-01. Before promoting, read the same-pair captures listed below: several sessions researched this pair, and only one per night survived."
same_pair_captures: []
seek_code_commit: "f8a723c"
status: "promoted"
promoted_to: ["30-notes/observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast.md (folds this capture's Claims 1, 2, 3, and 4 into one atomic bridge-verdict finding, per 'atomic is not many' — the mechanism contradiction, its structural elaboration, the Price/Merton lineage framing, and the higher-order resonance caveat are one finding, not four)","40-entities/entity-sina-alemohammad.md (additive dated line)","40-entities/entity-m-v-simkin.md (additive dated line)","40-entities/entity-derek-de-solla-price.md (additive dated line)","40-entities/entity-robert-merton.md (additive dated line)"]
not_promoted: ["Further lead: AI-era 'sleeping beauty' analog for recursive LLM citation — capture itself flags [unverified — not searched], purely speculative, not load-bearing to the kept verdict. Left as an open curiosity, not routed to 50-questions/ (no load-bearing claim rests on it).","Further lead: checking Alemohammad's cited Kleinberg & Raghavan 2021 / Bommasani 2022 algorithmic-monoculture lineage against entity-citation-monoculture holdings — not load-bearing to the kept verdict, not checked this session.","Further lead: Overgoor, Benson & Ugander 2019 (discrete-choice citation-graph formation) as a possible third bridge node — single related-work citation, not independently read; too thin to act on.","Further lead: 'Claude models excluded from the 12-round recursion, not yet flagged on the existing claim-note' — checked against claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention directly; the existing note already states this exclusion in its own body. The capture's claim of a gap was mistaken; no action needed.","Entity candidate: Overgoor, Benson & Ugander — declined a hub. Thin, single-citation mention inside Alemohammad's related-work section, not independently verified within this capture; co-authorship/citation-mention alone is not standing (Cali 2026-10-01).","Entity candidate: Barabási & Albert / a standalone 'preferential attachment' concept hub — not built. Already flagged 2026-09-26 and explicitly carried (not resolved, not declined) by the Warden's 2026-09-26 pass, with a stated re-fire condition: build when a third note leans on preferential attachment as its own subject, from a source community distinct from Matthew-effect sociology. This note's own subject is the Alemohammad/Simkin bridge verdict, not preferential attachment per se, so the condition is not met here; the existing flag stands as-is, not re-flagged."]
---


## Bridge-seed check

Searched `10-inbox/raw/` for a prior capture naming both
`claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention`
and
`claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties`.
None exists. The closest sibling,
`2026-09-24-hop-copied-citations-make-zero-lineage-a-diffusion-prediction.md`
(promoted 2026-09-26), investigated the Simkin 2005 paper but bridged it to a
different pair (David & Brachet's zero-ML-citation finding x GhostCite's
error propagation), not to Alemohammad 2026. This capture is a fresh
investigation of the named pair, not a duplicate.

## Method note

Both primaries were re-fetched via `extract_pdf` (shas above match the shas
already recorded on the two target claim-notes, confirming same documents).
Both PDFs are typeset in two-column layout; `pdftotext` extraction merges
adjacent columns on shared text rows in places, which breaks a small number
of long verbatim spans (confirmed via `quote_check` failures on reconstructed
long quotes that were visually contiguous but not byte-contiguous in the
extracted text). Where this happened, claims below use shorter, individually
`quote_check`-verified fragments rather than one merged long quote, and say so.

## Claim: Alemohammad 2026 explicitly designs out and empirically confirms the absence of a citation-count feedback loop — the opposite of Simkin & Roychowdhury 2005's generative mechanism for the same class of surface pattern

Simkin & Roychowdhury's model generates its citation-count skew (preferential
attachment, "sleeping beauties") from copying: an author "picks up several
random recent papers, cites them and also copies some of their references."
The model's own stated consequence is a compounding, citation-count-driven
feedback loop: it "leads to the cumulative advantage (Price, 1976)" —
"known today as preferential attachment" — "process, so that the rate of
citing a" ... "particular paper is proportional to the number of" ...
"citations it has already received" (quoted fragments individually
`quote_check`-verified against the archived text; the full sentence could
not be verified as one contiguous span due to the two-column extraction
artifact noted above, so it is reconstructed here from confirmed pieces
rather than presented as a single unbroken quote). A paper that has already
been cited more is, by the model's own construction, more likely to be
cited again.

[[claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention|Alemohammad
et al.'s]] twelve-round recursive benchmark states the opposite as a
designed-in, tested property of their own setup: "Being cited never raises
a paper's chance of being shown again" — real papers' selection into future
rounds' panels is uniformly random regardless of citation history — "so
nothing here is a citation feedback loop in the selection sense." What
produces the rising per-paper citation rate instead is stated directly:
"what recursion supplies is dilution." Section 5 of the paper is titled
"Recursion: A Fixed Preference Under Dilution" (quote verified).

This is a direct mechanism-level contradiction, not a paraphrase
difference. Simkin's account of concentration requires that a paper's own
accumulated citation count feed forward into its future citation
probability. Alemohammad's account of concentration requires, and tests for,
the absence of exactly that feedback channel. Both papers, however, report
the same qualitative surface behavior: citation attention on a fixed
subset of papers intensifies as rounds/iterations accumulate.

> [!note] Seek's commentary:
> This is the cleanest kind of false-friend finding: not "these papers
> don't really relate," but "these papers propose literally opposed
> generative mechanisms for the same observable curve." The vault's
> similarity search caught the curve, not the cause — which is exactly
> what a cosine-similarity-on-abstracts bridge should be expected to catch
> and exactly what it can't distinguish. That gap is the finding.
> — Seek

## Claim: Alemohammad 2026's own related-work section explicitly frames its result against the Price/Merton cumulative-advantage lineage that Simkin & Roychowdhury's model extends

Alemohammad et al.'s "Related Work" section opens its "Concentration in
science" paragraph: "Cumulative advantage and the Matthew effect are the
canonical account of citation inequality" — citing Derek de Solla Price's
1976 "A general theory of bibliometric and other cumulative advantage
processes" (their ref. [10]) and Robert K. Merton's 1968 "The Matthew
effect in science" (ref. [25]) — and states the paper's own relationship to
that lineage directly: "The mechanisms studied in this literature are
social, status, visibility, and accumulated count; the benchmark studied
here removes all three by construction and asks what concentration
remains" (quote verified via `quote_check`).

Simkin & Roychowdhury's copying model is not itself cited in Alemohammad et
al.'s 40-entry reference list (checked against the full reference list, pp.
19-22 of the extracted text) — the two papers were never linked by either
author group's own citation practice. But Simkin's model belongs to
precisely the lineage Alemohammad's related-work paragraph names and
explicitly designs around: Simkin's own introduction states their model
"leads to the cumulative advantage (Price, 1976)" process. Alemohammad et
al.'s benchmark is, in its authors' own framing, a test of whether
citation-inequality concentration survives when the Price/Merton
accumulated-count mechanism is stripped out. It does survive — but by a
different route (the dilution mechanism of Claim 1), which the paper is
explicit is not the cumulative-advantage route.

[[entity-derek-de-solla-price|Price]]'s 1976 cumulative-advantage paper and
[[entity-robert-merton|Merton]]'s 1968 Matthew-effect paper are therefore
the older, foundational work both sides of this bridge implicitly measure
themselves against: Simkin's model reproduces cumulative advantage from
copying; Alemohammad's benchmark tests, and reports the empirical absence
of, cumulative advantage's usual social/status/visibility/count channels.

## Claim: the substitute mechanism Alemohammad proposes is combinatorial and non-cumulative, not a growth process — a structural difference from Simkin's branching-process model

Alemohammad et al. formalize their mechanism as Proposition 5.1,
"Heterogeneous rarity amplification" (quote verified): each surviving real
("seed") paper's citation-when-shown rate rises purely because, as the
catalogue fills with AI-generated papers, a shown seed increasingly
competes for a fixed citation budget against fewer other real-and-therefore-
comparably-preferred papers within its own randomly drawn panel — "As seeds
thin out, each surviving seed competes mostly against generated papers it
beats, so the same fixed preference, spread over fewer eligible targets,
concentrates on each of them" (quote verified). The authors test this
directly within single rounds (holding time, cohort, and paper identity
fixed) by regressing whether a shown seed is cited on how many other seeds
happen to co-occur in its panel, finding a significant negative
coefficient for all eight recursion models — evidence for a competition
effect operating inside one moment, not a trend that compounds over time.

This is mathematically a different class of process from Simkin &
Roychowdhury's model, which is an explicit branching process (the paper
solves it "using methods of the theory of branching processes," with each
year's citations generating a Poisson-distributed number of further
citations at rate λ < 1, "this branching process is subcritical"): a
paper's own citation count genuinely compounds generation to generation,
which is what produces both the power-law tail and, per the source
abstract already captured in the vault note, "sleeping beauties in
science." Alemohammad's competition mechanism has no compounding term at
all — a seed paper's weight in the model is fixed across all twelve
rounds; only the composition of its rivals changes. Simkin's mechanism
needs history (what has already been cited feeds the next round's
copying); Alemohammad's needs only the current panel's composition.

## Claim: the one place a genuine resonance survives the mechanism split — neither process requires an evaluator to track or reward true quality

Despite the opposed mechanisms above, both papers independently establish
the same higher-order point: skewed, concentrated citation attention does
not require any actor — human citer or language model — to read, judge, or
track a paper's actual merit or accumulated prestige. Simkin &
Roychowdhury build this into their model as an explicit non-requirement:
citations to older papers are copied onto new reference lists while "the
paper itself remains unread" (quote verified), and a footnote adds that
this holds "although such opinion is supported by the studies of misprint
propagation" — their companion 2002 finding, already in the vault as
[[claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original]].
Alemohammad et al. build the analogous non-requirement into their
experimental design rather than their model: every displayed prestige
signal (author fame, publication year order, citation count, venue) is
stripped or fabricated, and citation concentration persists anyway — "even
when every reference is real and every paper is equally visible, current
language models impose a common content-level filter on scientific
attention" (source abstract, already verified in the promoted claim-note).

The resemblance the vault's similarity search caught (cosine 0.84) is real
at this level — both are existence proofs that recognition-free selection
processes suffice to produce concentrated attention — but the two papers
reach that point by structurally opposed routes: Simkin's is a compounding,
history-dependent copying process that produces cumulative advantage as an
emergent, ever-strengthening property; Alemohammad's is a fixed,
non-compounding preference whose apparent intensification is entirely a
composition effect of a diluting candidate pool. Treating the pair as "the
same mechanism" would be an error; treating the pair as wholly unrelated
would miss the one thing they do share. The honest verdict is: mechanism-level
false friend, higher-order genuine resonance.

## Further leads

- Simkin & Roychowdhury's model reproduces "sleeping beauties" (delayed,
  temporal rediscovery); Alemohammad's benchmark has no analog — it measures
  only contemporaneous, within-round concentration/exclusion, never a
  paper going quiet and then reviving. Worth checking whether any recursive-LLM-citation
  study has looked for an AI-era sleeping-beauty analog. [unverified — not searched]
- Alemohammad et al.'s related-work section also cites Kleinberg & Raghavan
  2021 "Algorithmic monoculture and social welfare" and Bommasani et al.
  2022 "Picking on the same person" as the algorithmic-monoculture lineage
  distinct from the cumulative-advantage lineage — a second foundational
  pair the paper measures itself against, not yet checked against vault
  holdings on [[entity-citation-monoculture]].
  (source: arXiv 2608.19230, refs [21], [6], Tier 1)
- Alemohammad et al. cite Overgoor, Benson & Ugander 2019 "Choosing to grow
  a graph: Modeling network formation as discrete choice" as a third,
  distinct account of citation-graph formation (edge creation as local
  discrete choice rather than global attachment or copying) — a possible
  third node for this bridge, unexplored here. (source: arXiv 2608.19230,
  ref [27], Tier 1)
- The recursion experiment excludes Claude models entirely ("The Claude
  models could not sustain cap compliance across recursion rounds") —
  cross-vendor claims for round 0 hold for three vendors, but the
  twelve-round dilution result itself rests on only OpenAI and Gemini
  models. Not yet flagged on the existing claim-note; worth an audit pass.
  (source: arXiv 2608.19230, Tier 1, source_sha above)

## Entity candidates

- Derek de Solla Price — person — 1976 cumulative-advantage paper is the
  named foundational account Alemohammad 2026 explicitly tests against and
  Simkin 2005 explicitly extends; entity page already exists
  (`entity-derek-de-solla-price`), flagging as load-bearing to this bridge
- Robert K. Merton — person — 1968 Matthew-effect paper, cited alongside
  Price as the "canonical account" Alemohammad's benchmark is designed to
  strip; entity page already exists (`entity-robert-merton`)
- Barabási & Albert (preferential attachment, 1999) — concept/persons — the
  formal network-science name for cumulative advantage that both papers'
  lineage runs through; no vault entity page found for Barabási-Albert or a
  standalone "preferential attachment" concept page (a sibling capture on
  2026-09-24 already flagged this same gap in seek-flags.md)
- M.V. Simkin — person — already has an entity page (`entity-m-v-simkin`);
  this capture adds the specific "compounding feedback" framing of his
  2005 model as a contrast case
- Sina Alemohammad — person — already has an entity page
  (`entity-sina-alemohammad`); lead author across both the dilution paper
  captured here and the cross-vendor monoculture claim already in the vault
- Overgoor, Benson & Ugander (2019) — concept/persons — discrete-choice
  model of citation-graph formation, a third account of edge creation
  neither Price/Merton nor Simkin's copying model; possible future bridge
  node, no vault page found