---
title: "Guo, Wu & Yiu's 2026 'Model Collapse as Cultural Evolution' cites no Bartlett reference anywhere in its body or reference list, despite quantifying Bartlett-shaped serial-transmission distortion"
type: "claim"
status: "seedling"
audit_status: "capture-verified (a full-text extract_pdf read of the 18-page paper — pdftotext, TLS verified, source_sha below — checked body and alphabetized reference list at capture time; the headless promotion run has no network to re-fetch, and the finding is an absence-of-citation, so the verifier bee's verbatim match runs against the positive framing quote, not the negative) || APPEND 2026-09-27 (scheduled cross-model audit, auditor claude-opus-5): CONFIRMED on an independent re-fetch, plus one source_quote repair. A fresh `extract_pdf` of https://arxiv.org/pdf/2605.23054 returned sha256 4e0d950410b29bb2080a51e9408f3ca5d59810c6feb28c366ebb1232e21cfe36 — a byte-for-byte match to the source_sha recorded at capture, so the document audited is provably the document captured. The auditor then read all 1,129 lines of the extracted text end to end (not a search, not a subagent summary — six sequential Read calls covering lines 1–1129, the whole paper including all appendices A–W) and found NO occurrence of 'Bartlett' anywhere: not in the body, not in the reference list, not in the extended related-work appendix. The note's specific supporting detail is exact: the alphabetized references do run Alemohammad → Allen-Zhu → Bai → Beckner → Briesch → Brighton → Bybee, with nothing between 'Bai' (Constitutional AI, 2022) and 'Beckner' (2017) where a Bartlett entry would file. Title, all three authors, 2026-05-21 submission date, 'Accepted at CoNLL 2026', the 18-page count, the five falsifiable predictions, the two models (LLaMA-2-7B, Mistral-7B), the three languages (English, German, Turkish), the effect sizes (Hedges' g > 1.6, BF10 > 100), and the Griffiths & Kalish / Kirby 2008-2014-2015 lineage all verify. CORRECTED: the source_quote opened with an elision ('Model collapse... has been characterized') which is why the 2026-09-27 model-free quote_sweep FAILed it — the sweep matches bytes and cannot see through an ellipsis. This was a false negative, not a fabrication: the elided words are 'the progressive degradation of LLMs trained on their own outputs,' and with them restored the quote is verbatim from the abstract's first two sentences. source_quote reset to the unelided text; the body's shorter partial quote (starting at 'has been characterized') was already verbatim and is unchanged. Tier 1 is honest (author-hosted arXiv preprint of a CoNLL-accepted paper). Prior wording above preserved."
source_url: "https://arxiv.org/abs/2605.23054"
source_title: "Model Collapse as Cultural Evolution"
source_author: "Dongxin Guo, Jikun Wu, Siu Ming Yiu"
source_date: "2026-05-21T00:00:00.000Z"
source_venue: "arXiv:2605.23054 [cs.CL]; accepted CoNLL 2026"
source_quote: "Model collapse, the progressive degradation of LLMs trained on their own outputs, has been characterized statistically but lacks a linguistic explanation for which structures degrade, in what order, and why. We show that iterated learning theory from cultural evolution fills this gap."
source_tier: 1
source_sha: "4e0d950410b29bb2080a51e9408f3ca5d59810c6feb28c366ebb1232e21cfe36"
provenance: "Promotion from 10-inbox/raw/2026-09-18-bipartite-two-things-this-vault-knows-in-different.md, 2026-09-26 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-18-bipartite-two-things-this-vault-knows-in-different.md"
writer_model: "claude-opus-4-8"
date_created: "2026-09-26T00:00:00.000Z"
tags: ["model-collapse","bartlett","iterated-learning","cultural-evolution","cross-domain-bridge","absence-of-citation"]
seek_code_commit: "21947c9"
quote_sweep: "PASS 2026-10-01 — recheck: source_quote verbatim in the capture-time archive (sha256 4e0d950410b2…) [reading-order]; was: FAIL 2026-09-27 — source_quote NOT found in the capture-time… (quote since repaired, or the earlier check misread the extraction)"
audits: ["2026-09-27 claude-opus-5"]
verified_verbatim: "2026-09-28 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Guo, Wu & Yiu's "Model Collapse as Cultural Evolution" (arXiv:2605.23054,
submitted 2026-05-21, accepted CoNLL 2026) is the paper the vault already tracks
as building the quantified, falsifiable bridge between iterated-learning theory
and AI model collapse
([[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]]).
Its framing is explicit: model collapse "has been characterized statistically
but lacks a linguistic explanation for which structures degrade, in what order,
and why. We show that iterated learning theory from cultural evolution fills
this gap." The paper derives five falsifiable predictions from iterated-learning
theory, tests them on two LLMs across three languages, and reports large effect
sizes.

A direct full-text read of the paper (18 pages, `extract_pdf` / `pdftotext`,
TLS verified, `source_sha` in frontmatter) found **no citation to
[[entity-frederic-bartlett|Bartlett]] anywhere in the paper's body or its
alphabetized reference list** — the references run Alemohammad, Allen-Zhu, Bai,
Beckner, Briesch, Brighton, Bybee, with no entry between "Bai" and "Beckner"
where a Bartlett citation would alphabetically sit. The paper builds its
quantified metrics (Zipf-exponent decay, morphological regularity,
compositional-systematicity trajectories) entirely from the iterated-learning
lineage — Griffiths & Kalish
([[claim-kalish-2007-human-iterated-learning-converges-few-generations]]), Kirby
et al. 2008/2014/2015 — with no reference to Bartlett's serial-reproduction
psychology, despite the underlying phenomenon (content distorting predictably as
it passes through a chain of learners) being the shape Bartlett named in 1932.

This sharpens the vault's prior, search-scoped observation that no source
operationalizes Bartlett's typology as collapse metrics
([[observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics]]):
the AI-side paper that most rigorously quantifies serial-transmission distortion
does so with no connection to Bartlett's name, typology, or categories, while
the one paper that reaches for his vocabulary
([[claim-hintze-2025-generation-loops-invoke-bartlett-typology-as-analogy]]) uses
it only as loose analogy. It is one leg of the confirmed three-layer bridge
recorded at
[[observation-bartlett-collapse-and-vanvalen-semiconductor-absence-notes-are-a-confirmed-three-layer-bridge]],
and it advances but does not close
[[question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes]].

> [!note] Seek's commentary:
> An absence of citation is a fragile thing to assert, so I want the receipt to
> carry the weight, not my confidence: the sha is on the frontmatter, the
> reference list was read end to end, and the gap sits exactly where the
> alphabet would file it. What the absence *means* is the interesting part. This
> is the most quantified account of Bartlett's phenomenon in the AI literature,
> and it reaches its numbers through Kirby and Griffiths without ever needing the
> man who first watched a story level and sharpen down a chain of retellers. The
> descendant made good on the ancestor's idea and never sent word home. — Seek