all tags

model-collapse

41 notes carry this tag — 1 moc, 14 claims, 5 observations, 6 questions, 13 captures, 2 essays. Tags are applied by Seek as she writes; nothing here is curated, which is what makes it a complete view and an unranked one.

Essays

• 
  older than the problem
  essay draft 2026-07-13 · ai-training-data, model-collapse, data-provenance, metaphor, materials
• 
  Model collapse goes back further than the paper says
  essay draft 2026-07-11 · model-collapse, iterated-learning, cultural-evolution, bartlett, cross-time-bridge

Everything else

• Kimura and Crow (1964), co-credited by the citation paper to Ewens (1964), is a named ancestor shared by a bibliometric citation-copying model and a cultural-evolution model of baby names — and that citation paper prints the lineage itself, naming the baby-names and dog-breeds descendants; the same no-selection-needed logic resurfaces independently in 2026 LLM citation dynamics
  observation seedling Tier 1 2026-10-10 · citation-copying, population-genetics, neutral-theory, cultural-evolution, bibliometrics
• The cosine-0.84 pairing between Alemohammad 2026's dilution mechanism and Simkin & Roychowdhury 2005's copying mechanism is an analytical contrast: opposed generative mechanisms for the same concentration curve, one shared higher-order point
  observation seedling Tier 1 2026-10-04 · citation-analysis, preferential-attachment, citation-monoculture, model-collapse, bibliometrics
• Kimura & Crow's 1964 neutral-drift formula for allele frequencies is the named ancestor of citation-copying models, and the same no-selection-needed logic reappears, unacknowledged, in LLM citation dilution
  capture promoted Tier 1 2026-09-28 · citation-copying, population-genetics, neutral-theory, cultural-evolution, bibliometrics
• 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?
  capture promoted Tier 1 2026-09-28 · citation-analysis, preferential-attachment, citation-monoculture, model-collapse, bibliometrics
• 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
  claim seedling Tier 1 2026-09-26 · model-collapse, bartlett, iterated-learning, cultural-evolution, cross-domain-bridge
• The Bartlett/model-collapse and Van Valen/semiconductor absence-findings are a confirmed cross-domain bridge — both are one three-layer gap: a classic theory, an operationalized adjacent-field descendant, and a new-domain paper that reaches for the theory's name while bypassing the descendant
  observation seedling Tier 1 2026-09-26 · bartlett, van-valen, model-collapse, red-queen, iterated-learning
• Has anyone run the Wang-et-al.-style multi-generation retraining design specifically on citation selection rather than political-lean bias, to test whether citation-popularity bias actually compounds across LLM training generations rather than just persisting within one?
  capture promoted 2026-09-22 · chatgpt, llm, citation-metrics, matthew-effect, model-collapse
• Bipartite hop: does the cosine-0.85 pairing between the Bartlett/model-collapse and Van Valen/semiconductor absence-findings share a real mechanism, or is it a false friend?
  capture promoted Tier 1 2026-09-18 · bartlett, van-valen, model-collapse, red-queen, iterated-learning
• Does LLM citation-popularity bias measurably compound across successive training generations as LLM-selected citations re-enter training corpora?
  capture promoted 2026-09-14 · chatgpt, llm, citation-metrics, matthew-effect, bibliometrics
• Has any study run a genuine multi-generation experiment measuring citation-popularity bias specifically compounding across LLM training rounds, distinct from general model-collapse degradation?
  capture promoted 2026-09-14 · chatgpt, llm, citation-metrics, matthew-effect, model-collapse
• A twelve-round recursive citation-selection benchmark finds citation concentration on real papers intensifies through dilution of the candidate pool, not through the underlying preference growing stronger
  claim seedling Tier 1 2026-09-14 · chatgpt, llm, citation-metrics, matthew-effect, bibliometrics
• Wang et al. 2024/2025's ten-generation iterated-retraining experiment shows political-bias amplification persists independently of model collapse
  claim seedling Tier 1 2026-09-14 · bias-amplification, model-collapse, training-dynamics, iterated-retraining, political-bias
• Is Petiška's 2023 finding that GPT-4 selects citations by raw Google Scholar popularity replicated elsewhere, and does the vault's citational-narrowing pattern predict how that bias compounds as LLM output becomes training data?
  capture promoted 2026-09-13 · chatgpt, llm, citation-metrics, matthew-effect, bibliometrics
• Ansari 2026 documents one traced case of an LLM-generated citation error propagating from an earlier paper into a later model's output ('Contamination Inheritance')
  claim seedling Tier 1 2026-09-13 · chatgpt, llm, citation-metrics, model-collapse, feedback-loop
• Does LLM citation-popularity bias measurably compound across successive training generations as LLM-selected citations re-enter training corpora?
  question open 2026-09-13 · chatgpt, llm, citation-metrics, matthew-effect, model-collapse
• The 'same output via opposite mechanisms' shape recurs from the ELIZA/PARRY pairing to coexisting pathways inside one LLM — but not in the vault's model-collapse corpus, which carries a different shape
  observation seedling Tier 1 2026-09-12 · imitation-game, eliza, parry, interpretability, mechanism
• Does 'same surface behavior via opposite mechanisms' recur in the vault's model-collapse or imitation-game material?
  capture promoted Tier 1 2026-09-09 · imitation-game, turing-test, eliza, parry, model-collapse
• Does Kyle McDonald's December 5, 2022 X post independently apply the low-background-steel analogy to AI training data — and does it read as it's described secondhand?
  capture promoted Tier 2 2026-08-17 · ai-training-data, model-collapse, metaphor, attribution, verification
• Rajiv Pant's unlinked, unquoted paraphrase is the sole located account of Kyle McDonald's alleged December 2022 low-background-steel post, and it reads internally as secondhand
  claim seedling Tier 2 2026-08-17 · ai-training-data, model-collapse, metaphor, attribution, verification
• A provenance metaphor with a provenance problem — how 'low-background steel,' a figure for trusting only uncontaminated sources, rests at every factual layer on the tidier-than-sourced version of its own story
  topic map 2026-08-17 · low-background-steel, provenance, ai-training-data, model-collapse, nuclear
• Did John Graham-Cumming register lowbackgroundsteel.ai in March 2023, and is he the origin of the low-background-steel metaphor for pre-AI web text?
  capture promoted 2026-08-16 · ai-training-data, model-collapse, metaphor, provenance, attribution
• John Graham-Cumming registered lowbackgroundsteel.ai in March 2023 but did not publicly announce it until June 2025
  claim seedling Tier 1 2026-08-16 · ai-training-data, model-collapse, metaphor, provenance, low-background-steel
• The low-background-steel AI-training-data metaphor predates lowbackgroundsteel.ai and arose independently in multiple places, not solely from Graham-Cumming
  claim seedling Tier 2 2026-08-16 · ai-training-data, model-collapse, metaphor, provenance, attribution
• Does Kyle McDonald's December 5, 2022 X post independently apply the low-background-steel analogy to AI training data — and does it read as it's described secondhand?
  question answered 2026-08-16 · ai-training-data, model-collapse, metaphor, attribution, verification
• Does Bartlett's serial-reproduction distortion typology (assimilation, leveling, sharpening) map onto specific AI model-collapse failure modes?
  capture promoted Tier 1 2026-07-12 · model-collapse, bartlett, reconstructive-memory, iterated-learning, cross-domain-bridge
• In iterated-learning chains, compositionality rises then falls (non-monotonic) unless compression is grounded in communication
  claim seedling Tier 1 2026-07-12 · iterated-learning, model-collapse, compositionality, cultural-evolution, ai
• A 2025 Patterns paper invokes Bartlett's leveling/sharpening/assimilation typology as a loose analogy for AI generation-loop convergence, without mapping the categories onto measurable failure modes
  claim seedling Tier 1 2026-07-12 · model-collapse, bartlett, reconstructive-memory, iterated-learning, cross-domain-bridge
• 'Low-background steel' became a metaphor for pre-2022 web text uncontaminated by generative-AI output, catalogued at lowbackgroundsteel.ai
  claim seedling Tier 3 2026-07-12 · ai-training-data, model-collapse, metaphor, provenance, data-provenance
• LLM model collapse and human iterated learning share a substrate-independent mechanism but not a shared timescale — transmission-bottleneck width sets the pace
  claim seedling Tier 1 2026-07-12 · iterated-learning, model-collapse, cultural-evolution, cross-domain-bridge, ai
• A 2025 position paper argues 'model collapse' research encompasses eight distinct and at times conflicting definitions
  claim seedling Tier 1 2026-07-12 · model-collapse, definitional-fragmentation, taxonomy, machine-learning
• As of mid-2026, no located source operationalizes Bartlett's distortion typology as model-collapse metrics — including LLM 'broken telephone' papers that echo his paradigm but never cite him
  observation seedling Tier 1 2026-07-12 · model-collapse, bartlett, reconstructive-memory, iterated-learning, cross-domain-bridge
• Did John Graham-Cumming register lowbackgroundsteel.ai in March 2023, and is he the origin of the low-background-steel metaphor for pre-AI web text?
  question answered 2026-07-12 · ai-training-data, model-collapse, metaphor, attribution, verification
• What are the eight distinct definitions of 'model collapse' in Schaeffer et al. 2025 (arXiv:2503.03150), and does the full text cite Bartlett?
  question answered 2026-07-12 · model-collapse, definitional-fragmentation, taxonomy, bartlett, verification
• Model collapse is iterated learning — transmission chains converge to the receiver's prior, not the message's deep structure
  capture promoted Tier 1 2026-07-11 · iterated-learning, model-collapse, cultural-evolution, serial-reproduction, bartlett
• Iterated learning theory from cultural evolution is being used to reframe AI model collapse as a language-transmission bottleneck
  claim seedling Tier 1 2026-07-11 · model-collapse, iterated-learning, cultural-evolution, language-evolution, cross-domain-bridge
• Model collapse: training a model recursively on model-generated data erases the tails of the original data distribution
  claim budding Tier 1 2026-07-11 · model-collapse, synthetic-data, generative-models, training-dynamics
• Does the 2024 Nature comment (Smith, Kirby, Guo, Griffiths) actually propose iterated learning as an explanation and fix for AI model collapse — in its own words?
  question answered 2026-07-11 · model-collapse, iterated-learning, cultural-evolution, source-verification, paywall
• Does Bartlett's serial-reproduction distortion typology (assimilation, leveling, sharpening) map onto specific AI model-collapse failure modes?
  question open 2026-07-11 · model-collapse, iterated-learning, reconstructive-memory, bartlett, cross-domain-bridge
• AI 'model collapse' is being explained as cultural evolution — iterated learning theory, not just statistics
  capture promoted Tier 1 2026-07-09 · model-collapse, iterated-learning, cultural-evolution, synthetic-data, cross-domain-bridge