Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

This paper claims that an ecosystem of a few diverse language models, each fine-tuned on a slice of the same fixed dataset and then retrained on their pooled output, preserves the original data distribution better than a single model or a v

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 15:53 UTC pith:ODR4PEDT

load-bearing objection The M=4 optimum is visible, but the abstract promises results the experiments never deliver — worth refereeing with major revision, not rejecting out of hand. the 4 major comments →

arxiv 2512.15011 v3 pith:ODR4PEDT submitted 2025-12-17 cs.LG cs.AIcs.CYcs.MA

Epistemic diversity across language models mitigates knowledge collapse

Damian Hodel , Jevin D. West This is my paper
classification cs.LG cs.AIcs.CYcs.MA
keywords model collapseknowledge collapseepistemic diversityAI ecosystemmodel pluralismself-trainingperplexityHill-Shannon Diversity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether concentrating AI knowledge in a few dominant models (monoculture) or spreading it across many diverse models better resists 'knowledge collapse'—the degradation that occurs when models are retrained on AI-generated data. It finds that diversity helps, but only up to a point: under a fixed data and compute budget, an ecosystem of four models keeps perplexity on the original test set lowest over ten self-training iterations, for both GPT-2 and OPT-125m. A single model decays quickly; sixteen models start poorly because each gets too little data. The authors interpret the inverted-U as a trade-off between ensemble expressivity (diverse models can represent a richer mixture) and each model's approximation capacity on the true distribution. If right, the result argues for monitoring and actively cultivating plural, community-specific AI models rather than relying on one universal model.

Core claim

Under a fixed per-ecosystem resource budget, the paper reports that dividing a fixed training set among M=4 independently fine-tuned language models and retraining them on their collectively generated, shuffled, and redistributed output over ten iterations yields the lowest aggregated perplexity on the original test set, for both GPT-2 and OPT-125m. Low diversity (M=1,2) performs well early but degrades rapidly; high diversity (M=16) starts with high perplexity but stabilizes, even showing a negative perplexity trend. The optimal diversity is therefore greater than zero but not maximal, and the paper additionally claims that this optimal level rises with the number of self-training iteration

What carries the argument

Ecosystem diversity is instantiated by randomly segmenting the fixed training corpus into M disjoint subsets—one per model—while keeping the total number of tokens N constant, so each model fine-tunes on n=N/M tokens. At each iteration the M generated datasets are concatenated, shuffled, and uniformly redistributed, so all models train on the same mixture while generating from their own fixed prompt subsets. Diversity is quantified by the Hill–Shannon Diversity index, which here simplifies to the number of models D=M. The performance measure is mean perplexity on the original test set, and the theoretical frame is the three-error decomposition of model collapse (statistical approximation, fu

Load-bearing premise

The load-bearing premise is that randomly splitting one fixed dataset into disjoint prompt subsets creates persistent epistemic diversity between models, even though after the first round all models are retrained on the same pooled mixture and differ only in the fixed prompts they generate from; if that does not produce genuine differences in what the models know, the causal role of 'diversity' is not established.

What would settle it

Run the M=4 condition with all models sharing an identical generation prompt subset (no persistent difference) while keeping segmentation and pooling the same; if the benefit over M=1 vanishes, the effect is due to prompt identity rather than diversity. Alternatively, compare the current pooled-redistribution scheme against a scheme where each model trains only on its own lineage; identical collapse curves would show redistribution is irrelevant.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single dominant model is the worst long-term configuration: monoculture accelerates collapse once models train on their own outputs.
  • There is a sweet spot: a handful of differently trained models can outperform both one model and many models at the same compute/data budget.
  • The optimal number of models grows with continued self-training, implying that ecosystems need to become more diverse over time to stay stable.
  • Scaling up models and datasets amplifies collapse in homogeneous ecosystems, increasing the practical value of diversity.
  • For policy, this supports incentives for domain- and community-specific models and for ecosystem-level diversity monitoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The authors' setup instantiates diversity only through data segmentation; real systems could diversify along architecture, alignment, and interface dimensions, which might shift the optimum and amplify the effect, so the M=4 result may be a lower bound on real-world diversity benefits.
  • Because after the first iteration every model trains on the same pooled mixture, the persistent difference between models is only their fixed generation prompts; a natural extension is to measure the actual divergence between models' output distributions over time to verify that diversity persists and drives the effect.
  • The monotonic rise of optimal diversity with iterations could imply a 'diversity scaling law'—if recursion continues indefinitely, ever more specialized models are needed, which has implications for how to allocate a fixed data budget.
  • The paper's distinction between segmenting by reduction (splitting existing internet data) and by expansion (having models train on distinct new data) suggests a policy-relevant question: real ecosystems may need expansion-based diversity to achieve the same benefit for large models already trained on most available text.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper studies whether distributing a fixed amount of training data across an increasing number of language models (an "ecosystem") mitigates model collapse during recursive self-training. The authors fine-tune identical copies of GPT2 and OPT-125m on disjoint subsets of Wikitext2, then over 10 iterations generate new training data from each model's fixed generation subset, pool, shuffle, and redistribute the data uniformly. They report an optimal ecosystem diversity of D=M=4 (where D is the Hill-Shannon Diversity, which equals the number of models M by equal weights), giving the lowest 10-iteration average perplexity on the original test set. They interpret this as a trade-off between increased ensemble expressivity and reduced per-model approximation capacity.

Significance. The question is relevant and timely. If the result holds, it would provide a concrete, controlled demonstration that a diverse ecosystem of models can outperform a monoculture under a fixed data/compute budget, with implications for AI governance and data-sharing. The experimental design is simple and reproducible in principle, and the paper clearly states its assumptions. However, the present manuscript does not provide the evidence needed to support its central claims: the abstract goes beyond the reported results, the diversity manipulation is confounded, and the statistics are insufficient.

major comments (4)
  1. [Abstract and §4, Fig. 3] The abstract claims that the optimal diversity level increases monotonically with the number of self-training iterations, and that the effect is robust to mixing human- and model-generated data and to temperature sampling methods. None of these results appear in the body. Section 4 and Figure 3 only report a single aggregated optimum (D=M=4) over ten iterations for two model families (GPT2, OPT-125m). Either the experiments must be added or the abstract claims must be removed.
  2. [§3, Eq. (2) and Fig. 2] The manipulation D=M confounds the number of models with per-model sample size n=N/M. With a fixed total token budget N, increasing M reduces each model's training data, so the observed M=4 optimum could be a sample-size/regularization effect rather than a diversity effect. No control experiment isolates diversity from sample size, and no direct measure of epistemic diversity among models (e.g., output distribution divergence) is provided. The HSD reduces to M by the equal-weight definition, not by measuring any actual divergence. The causal claim that diversity per se mitigates collapse is therefore not established.
  3. [§4, Figs. 3-4] No repeated-seed or independent-run error bars are reported. The shaded region in Figure 4 is the combined standard deviation across models and data points within a single run, not across random seeds. With one trajectory per condition, the M=4 optimum may be noise. Multiple seeds and confidence intervals (or significance tests) are required to support the central empirical claim.
  4. [§3, Fig. 2] After the first iteration, the M generated datasets are concatenated, shuffled, and uniformly redistributed to all models, so every model trains on a random shard of the same pooled mixture. The only persistent difference between models is each model's fixed generation prompt subset. This is a weak instantiation of the stated 'epistemic diversity' and does not guarantee genuinely diverse learned distributions. The paper should either provide evidence that the models are actually diverse (e.g., measure divergence between their output distributions) or acknowledge that the effect could be driven by random partitioning or prompt sampling rather than diversity.
minor comments (4)
  1. [§6] Typo: 'larg-scale' should be 'large-scale'.
  2. [Fig. 1] The symbol 'm' in Figure 1 is not defined in the caption; it presumably denotes a model, but this should be stated.
  3. [§3.2] The notation n=N/M is used before it is defined; the definition appears later in the same paragraph. Please define n explicitly at first use.
  4. [§3.2] The dataset sizes are given as '7,500, 1,900, and 2,200 data points' after stating 'two-fifths of the original size (=N)'. Clarify whether N=7,500 and whether these counts refer to blocks of 128 tokens or something else.

Circularity Check

0 steps flagged

No circularity; the central empirical claim is independent of its inputs.

full rationale

The paper's derivation is self-contained and empirically grounded. The central quantity, aggregated perplexity on a fixed Wikitext2 test split (Eqs. 3–4), is measured, not fitted from the diversity variable. The Hill-Shannon Diversity reduces to M only because of equal model weights (Eq. 2), which is an operational labeling choice rather than a recycled conclusion. The optimality of M=4 is read directly from Figure 3/4 and is not assumed by the setup. Prior-work citations (e.g., Shumailov et al. [39]) supply methodology and background, not the paper's target result. The paper even flags the main confound explicitly ('The only difference is the per-model sample size (n=N/M)', Section 4), which weakens the causal interpretation of 'diversity' but does not make the empirical finding circular. The appendix and discussion also acknowledge limitations (e.g., footnote 6 about dataset opacity) without invoking them as evidence. No fitted parameter is renamed as a prediction, and no load-bearing argument rests on the authors' own prior work. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new entities (forces, particles, dimensions, conserved quantities) are introduced. The paper's epistemic framework is carried by borrowed concepts—knowledge collapse, AI monoculture, Hill-Shannon Diversity—and the load-bearing assumptions are about what perplexity measures and what random segmentation means for diversity.

free parameters (6)
  • number of models M grid = 1, 2, 4, 16
    Coarse, hand-chosen diversity levels; the optimum at M=4 could be an artifact of this grid.
  • training epochs = 5
    Hand-chosen following prior work; affects collapse dynamics.
  • learning rate = 1e-5
    Hand-chosen; follows Shumailov et al. and is not varied.
  • beam search width = 5
    Generation strategy follows prior work; affects contamination of synthetic data.
  • dataset subset fraction = 2/5 of Wikitext2
    Arbitrary reduction to limit compute; changes data budget and hence collapse dynamics.
  • per-model sample size n=N/M = N/M
    Determined by M, but is the true confound: diversity benefits are entangled with less data per model.
axioms (5)
  • domain assumption Model collapse framework of Shumailov et al. (statistical approximation, functional expressivity, functional approximation errors) transfers from single models to ecosystems.
    Section 5 builds the interpretation on these three error sources without derivation for the multi-model case.
  • domain assumption Perplexity on the original Wikitext2 test split measures alignment with the true distribution and hence knowledge preservation.
    Section 3.1 equates lower perplexity with better performance; this is a standard but limited proxy.
  • ad hoc to paper Random segmentation of fixed training data instantiates epistemic diversity.
    Section 3 assumes D=M; no direct measurement of output diversity is provided, and pooling+redistribution may erase specialization.
  • domain assumption Small models (GPT2, OPT-125m) and Wikitext2 are representative of real-world AI ecosystems.
    Section 6 acknowledges limitations, but the central claims generalize only under this assumption.
  • domain assumption Mean perplexity aggregated across models (Eqs. 3 and 4) is a valid ecosystem performance measure.
    Mean perplexity treats all models equally and can hide distributional collapse within the ecosystem.

pith-pipeline@v1.3.0-alltime-deepseek · 12809 in / 13275 out tokens · 133097 ms · 2026-08-03T15:53:37.135776+00:00 · methodology

0 comments
read the original abstract

Artificial intelligence (AI) increasingly generates the very content used to train future AI systems. This feedback loop can degrade model quality, reduce informational diversity, and ultimately drive knowledge collapse, i.e. a degradation to a narrow and inaccurate set of ideas. We ask: to mitigate collapse, is it better to concentrate the internet's knowledge into a handful of dominant models (referred to as an AI monoculture), or to distribute it across a diverse ecosystem of models? To study the effect of diversity on model performance, we randomly segment the fixed training data across an increasing number of language models and evaluate the resulting ecosystems of models over ten self-training iterations. Our results show that diversity improves long-term performance of models, while monoculture accelerates collapse. Specifically, we observe that the optimal diversity level (i.e., the level that maximizes performance) increases monotonically with the number of self-training iterations. The observed effect is robust across various experimental settings, including different model families, parameter sizes, mixing human- and model-generated data, and temperature sampling methods, demonstrating the significance of ecosystem diversity for mitigating collapse. Moreover, our experiments with increased model and dataset sizes indicate that scaling up the system can amplify collapse in homogeneous ecosystems, thereby increasing the diversity benefits. In the presence of AI monoculture, our results suggest considering information environments with specialized AI models that maintain and enhance diversity in knowledge production, akin to the benefits of ecological diversity in biology and social systems.

Figures

Figures reproduced from arXiv: 2512.15011 by Damian Hodel, Jevin D. West.

Figure 1
Figure 1. Figure 1: Illustration of model collapse in AI ecosystems. Because of a model’s (m) inherent biases (in learning and resampling) and finite data, it cannot perfectly regenerate the information space (e.g., concepts, claims, facts) it was trained on. Some information becomes distorted, and some is lost entirely due to low sampling probabilities. When this training–generation cycle is repeated, the model’s output beco… view at source ↗
Figure 2
Figure 2. Figure 2: Instantiating ecosystem diversity and evaluating model performance. We examine the perfor￾mance of models 𝑚 within collections of increasing diversity 𝐷 when trained on collective output across 10 training iterations 𝑡. The number of models 𝑀 equals the Hill-Shannon Diversity 𝐷, which is instantiated through segmentation of the collected output. At each iteration, we concatenate and shuffle the 𝑀 generated… view at source ↗
Figure 3
Figure 3. Figure 3: AI ecosystems exhibit an optimal degree of diversity (Hill–Shannon Diversity 𝐷) for perfor￾mance, operationalized as the mean perplexity on the original test set averaged across 10 training–generation iterations. For both model types, GPT2 and OPT-125m, this perplexity minimum occurs at 𝐷 = 𝑀 = 4. The colored symbols correspond to the colors in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Diversity across models 𝐷 mitigate model collapse. Comparison of ecosystems’ perplexity with increasing diversity (𝐷 = 𝑀) at fixed dataset size across ten iteration steps. For both model types, 𝑀 = 4 yields the lowest mean perplexity at the early and late stages of evolution. After 10 iterations, the perplexity increasing rate is positive for low diversity (𝑀 = 1, 2), near 0 for 𝑚 = 4, and negative for the… view at source ↗
Figure 5
Figure 5. Figure 5: Perplexity distributions widen with increasing diversity. Mixed perplexity distribution on the original Wikitext2 train data points, evaluated by all models in a given ecosystem with diversity (𝐷 = 𝑀 = 1, 2, 4, 16) at iteration steps 𝑡 = 0 (left) and 𝑡 = 9 (right). As ecosystems’ diversity increases, the distributions become wider indicating greater expressivity. Simultaneously, their means shift toward hi… view at source ↗
Figure 6
Figure 6. Figure 6: illustrates the probability distributions of the information space. Information Information Information True Distribution Training Data0 Evaluation Data Model Statistical approximation error Functional expressivity and functional approximation errors Density Density p p̂ θ (t=0) data (t=0) p̂ pdata data (t=1) Information Density p̂data (test) Model Output / Training Data1 Density Information Density N0 *N0… view at source ↗
Figure 7
Figure 7. Figure 7: shows the perplexity progress of the two model types across ten iteration steps, including the models’ perplexity for the off-the-shelf models. (a) OPT-125m (b) GPT2 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Diversifying Personalized Research Ideation against AI-Induced Homogenization

    cs.AI 2026-07 conditional novelty 5.5

    DivAlign cuts community nearest-neighbor idea similarity from 0.663 to 0.608 while keeping about 99.9% of researcher–direction fit versus independent top-choice selection.

  2. Position: Prioritize Identifying Structure, Not Complex Models, for Scientific Discovery

    stat.ML 2026-05 unverdicted novelty 4.0

    Mechanistic learning from ML is generically underdetermined in high-dimensional proxy regimes, with LLMs worsening the problem by collapsing many possible explanations into one fluent narrative.

Reference graph

Works this paper leans on

59 extracted references · 15 canonical work pages · cited by 2 Pith papers · 4 internal anchors

  1. [1]

    Baraniuk

    Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. 2023. Self-Consuming Generative Models Go MAD. https://doi.org/10.48550 /arXiv.2307.01850 arXiv:2307.01850 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.01850 2023

  2. [2]

    Kareem Amin, Sara Babakniya, Alex Bie, Weiwei Kong, Umar Syed, and Sergei Vassilvitskii. 2025. Escaping Collapse: The Strength of Weak Data for Large Language Model Training. https://doi.org/10.48550/arXiv.2502.08924 arXiv:2502.08924 [cs]

    work page doi:10.48550/arxiv.2502.08924 2025

  3. [3]

    Ang, and Nabil Abou Baker

    Daniela Anderson, Katie Lien, Chibueze Agwu, Phillip S. Ang, and Nabil Abou Baker. 2023. The Bias of Medicine in Sickle Cell Disease.Journal of General Internal Medicine38, 14 (Nov. 2023), 3247–3251. https://doi.org/10.1007/s11606- 023-08392-0

    doi:10.1007/s11606- 2023

  4. [4]

    Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan

    Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, K...

    Pith/arXiv arXiv 2022

  5. [5]

    Quentin Bertrand, Avishek Joey Bose, Alexandre Duplessis, Marco Jiralerspong, and Gauthier Gidel. 2024. On the Stability of Iterative Retraining of Generative Models on their own Data. https://doi.org/10.48550/arXiv.2310.00429 arXiv:2310.00429 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.00429 2024

  6. [6]

    Borhane Blili-Hamelin, Christopher Graziul, Leif Hancox-Li, Hananel Hazan, El-Mahdi El-Mhamdi, Avijit Ghosh, Katherine A Heller, Jacob Metcalf, Fabricio Murai, Eryk Salvaggio, Andrew J Smart, Todd Snider, Mariame Tighanimine, Talia Ringer, Margaret Mitchell, and Shiri Dori-Hacohen. 2025. Position: Stop treating ‘AGI’ as the north-star goal of AI research....

    2025

  7. [7]

    Creel, Ananya Kumar, Dan Jurafsky, and Percy S

    Rishi Bommasani, Kathleen A. Creel, Ananya Kumar, Dan Jurafsky, and Percy S. Liang. 2022. Picking on the same person: Does algorithmic monoculture lead to outcome homogenization?Advances in Neural Information Processing Systems35 (2022), 3663–3678. https://proceedings.neurips.cc/paper_files/paper/2022/hash/17a234c91f746d9625a75cf8a8731ee2- Abstract-Confer...

    2022

  8. [8]

    Elvis Dohmatob, Yunzhen Feng, Arjun Subramonian, and Julia Kempe. 2024. Strong Model Collapse. https: //doi.org/10.48550/arXiv.2410.04840 arXiv:2410.04840 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.04840 2024

  9. [9]

    Elvis Dohmatob, Yunzhen Feng, Pu Yang, Francois Charton, and Julia Kempe. 2024. A Tale of Tails: Model Collapse as a Change of Scaling Laws. https://doi.org/10.48550/arXiv.2402.07043 arXiv:2402.07043 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.07043 2024

  10. [10]

    Heather E. Douglas. 2009.Science, Policy, and the Value-Free Ideal. University of Pittsburgh Pre. Google-Books-ID: LcFvKeOJRmgC

    2009

  11. [11]

    2019.PyTorch Lightning

    William Falcon and The PyTorch Lightning team. 2019.PyTorch Lightning. https://doi.org/10.5281/zenodo.3828935

    doi:10.5281/zenodo.3828935 2019

  12. [12]

    Sina Fazelpour and Daniel Steel. 2022. Diversity, Trust, and Conformity: A Simulation Study.Philosophy of Science89, 2 (April 2022), 209–231. https://doi.org/10.1017/psa.2021.25 Publisher: Cambridge University Press (CUP)

    work page doi:10.1017/psa.2021.25 2022

  13. [13]

    Shangbin Feng, Taylor Sorensen, Yuhan Liu, Jillian Fisher, Chan Young Park, Yejin Choi, and Yulia Tsvetkov. 2024. Mod- ular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration. http://arxiv.org/abs/2406.15951 arXiv:2406.15951 [cs]

    Pith/arXiv arXiv 2024

  14. [14]

    Yunzhen Feng, Elvis Dohmatob, Pu Yang, Francois Charton, and Julia Kempe. 2024. Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification. https://doi.org/10.48550/arXiv.2406.07515 arXiv:2406.07515 [cs]. 12 Damian Hodel and Jevin D. West

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.07515 2024

  15. [15]

    Political Neutrality in AI Is Impossible- But Here Is How to Approximate It

    Jillian Fisher, Ruth E. Appel, Chan Young Park, Yujin Potter, Liwei Jiang, Taylor Sorensen, Shangbin Feng, Yulia Tsvetkov, Margaret E. Roberts, Jennifer Pan, Dawn Song, and Yejin Choi. 2025. Political Neutrality in AI Is Impossible- But Here Is How to Approximate It. https://doi.org/10.48550/arXiv.2503.05728 arXiv:2503.05728 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.05728 2025

  16. [16]

    Shi Fu, Yingjie Wang, Yuzhu Chen, Xinmei Tian, and Dacheng Tao. 2025. A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops. https://doi.org/10.48550/arXiv.2502.18865 arXiv:2502.18865 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.18865 2025

  17. [17]

    Timnit Gebru and Émile P. Torres. 2024. The TESCREAL bundle: Eugenics and the promise of utopia through artificial general intelligence.First Monday(April 2024). https://doi.org/10.5210/fm.v29i4.13636

    doi:10.5210/fm.v29i4.13636 2024

  18. [18]

    Roberts, Diyi Yang, David L

    Matthias Gerstgrasser, Rylan Schaeffer, Apratim Dey, Rafael Rafailov, Henry Sleight, John Hughes, Tomasz Korbak, Rajashree Agrawal, Dhruv Pai, Andrey Gromov, Daniel A. Roberts, Diyi Yang, David L. Donoho, and Sanmi Koyejo

  19. [19]

    Mandeep Goyal and Qusay H. Mahmoud. 2024. A Systematic Review of Synthetic Data Generation Techniques Using Generative AI.Electronics13, 17 (Jan. 2024), 3509. https://doi.org/10.3390/electronics13173509 Publisher: Multidisciplinary Digital Publishing Institute

    doi:10.3390/electronics13173509 2024

  20. [20]

    David Herel and Tomas Mikolov. 2024. Collapse of Self-trained Language Models. http://arxiv.org/abs/2404.02305 arXiv:2404.02305 [cs]

    Pith/arXiv arXiv 2024

  21. [21]

    M. O. Hill. 1973. Diversity and Evenness: A Unifying Notation and Its Consequences.Ecology54, 2 (1973), 427–432. https://doi.org/10.2307/1934352 _eprint: https://esajournals.onlinelibrary.wiley.com/doi/pdf/10.2307/1934352

    doi:10.2307/1934352 1973

  22. [22]

    Shomik Jain, Vinith Suriyakumar, Kathleen Creel, and Ashia Wilson. 2024. Algorithmic Pluralism: A Structural Approach To Equal Opportunity. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’24). Association for Computing Machinery, New York, NY, USA, 197–206. https://doi.org/10.1145/3630106.36 58899

    work page doi:10.1145/3630106.36 2024

  23. [23]

    Lou Jost. 2006. Entropy and diversity.Oikos113, 2 (2006), 363–375. https://doi.org/10.1111/j.2006.0030-1299.14714.x _eprint: https://nsojournals.onlinelibrary.wiley.com/doi/pdf/10.1111/j.2006.0030-1299.14714.x

    arXiv 2006

  24. [24]

    Jon Kleinberg and Manish Raghavan. 2021. Algorithmic monoculture and social welfare.Proceedings of the National Academy of Sciences118, 22 (June 2021), e2018340118. https://doi.org/10.1073/pnas.2018340118

    doi:10.1073/pnas.2018340118 2021

  25. [25]

    Seth Lazar and Alondra Nelson. 2023. AI safety on whose terms?Science381, 6654 (July 2023), 138–138. https: //doi.org/10.1126/science.adi8982

    work page doi:10.1126/science.adi8982 2023

  26. [26]

    Manning, and James Zou

    Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Lepp, Wenlong Ji, Xuandong Zhao, Hancheng Cao, Sheng Liu, Siyu He, Zhi Huang, Diyi Yang, Christopher Potts, Christopher D. Manning, and James Zou. 2025. Quantifying large language model usage in scientific papers.Nature Human Behaviour(Aug. 2025), 1–11. https://doi.org/10.1038/s41562- 025-02273-8 Publisher: ...

    doi:10.1038/s41562- 2025

  27. [27]

    1990.Science as social knowledge: Values and objectivity in scientific inquiry

    Helen E Longino. 1990.Science as social knowledge: Values and objectivity in scientific inquiry. Princeton university press

    1990

  28. [28]

    Christina Lu and Max Van Kleek. 2024. Model Plurality: A Taxonomy for Pluralistic AI. (Oct. 2024). https: //openreview.net/forum?id=kil2mabTqx

    2024

  29. [29]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models.arXiv preprint arXiv:1609.07843(2016)

    Pith/arXiv arXiv 2016

  30. [30]

    Lisa Messeri and M. J. Crockett. 2024. Artificial intelligence and illusions of understanding in scientific research. Nature627, 8002 (March 2024), 49–58. https://doi.org/10.1038/s41586-024-07146-0 Publisher: Nature Publishing Group

    doi:10.1038/s41586-024-07146-0 2024

  31. [31]

    Tanisha Mishra, Edward Sutanto, Rini Rossanti, Nayana Pant, Anum Ashraf, Akshay Raut, Germaine Uwabareze, Ajayi Oluwatomiwa, and Bushra Zeeshan. 2024. Use of large language models as artificial intelligence tools in academic research and publishing among global clinical researchers.Scientific Reports14, 1 (Dec. 2024), 31672. https://doi.org/10.1038/s41598...

    work page doi:10.1038/s41598-024-81370-6 2024

  32. [32]

    Mozilla. 2024. Mozilla Report: How Common Crawl’s Data Infrastructure Shaped the Battle Royale over Generative AI. https://www.mozillafoundation.org/en/blog/Mozilla-Report-How-Common-Crawl-Data-Infrastructure-Shaped- the-Battle-Royale-over-Generative-AI/ Section: AI

    2024

  33. [33]

    Peterson

    Andrew J. Peterson. 2025. AI and the problem of knowledge collapse.AI & SOCIETY40, 5 (June 2025), 3249–3269. https://doi.org/10.1007/s00146-024-02173-x

    doi:10.1007/s00146-024-02173-x 2025

  34. [34]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. (2019)

    2019

  35. [35]

    Kristina Rolin. 2019. The Epistemic Significance of Diversity. InThe Routledge Handbook of Social Epistemology(1 ed.), Miranda Fricker, Peter J. Graham, David Henderson, and Nikolaj J. L. L. Pedersen (Eds.). Routledge, 158–166. https://doi.org/10.4324/9781315717937-16

    work page doi:10.4324/9781315717937-16 2019

  36. [36]

    Michael Roswell, Jonathan Dushoff, and Rachael Winfree. 2021. A conceptual guide to measuring species diversity.Oikos130, 3 (2021), 321–338. h t t p s : / / d o i . o r g / 1 0 . 1 1 1 1 / o i k . 0 7 2 0 2 _eprint: Epistemic diversity across language models mitigates knowledge collapse 13 https://nsojournals.onlinelibrary.wiley.com/doi/pdf/10.1111/oik.07202

    work page doi:10.1111/oik.07202 2021

  37. [37]

    Rylan Schaeffer, Joshua Kazdan, Alvan Caleb Arulandu, and Sanmi Koyejo. 2025. Position: Model Collapse Does Not Mean What You Think. https://doi.org/10.48550/arXiv.2503.03150 arXiv:2503.03150 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.03150 2025

  38. [38]

    Max Schaffelder and Albert Gatt. 2025. Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning. https://doi.org/10.48550/arXiv.2511.01490 arXiv:2511.01490 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.01490 2025

  39. [39]

    Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. 2024. AI models collapse when trained on recursively generated data.Nature631, 8022 (July 2024), 755–759. https://doi.org/10.1038/s4 1586-024-07566-y

    doi:10.1038/s4 2024

  40. [40]

    Miriam Solomon. 2006. Norms of Epistemic Diversity.Episteme3, 1-2 (June 2006), 23–36. https://doi.org/10.3366/epi. 2006.3.1-2.23

    work page doi:10.3366/epi 2006

  41. [41]

    Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S. Morcos. 2023. Beyond neural scaling laws: beating power law scaling via data pruning. https://doi.org/10.48550/arXiv.2206.14486 arXiv:2206.14486 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2206.14486 2023

  42. [42]

    Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan, Colin McDaniel, Alireza Ziabari, Jackson Trager, Ala Tak, Meng Chen, Fred Morstatter, and Morteza Dehghani. 2025. The Shrinking Landscape of Linguistic Diversity in the Age of Large Language Models. https://doi.org/10.48550/arXiv.2502.11266 arXiv:2502.11266 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.11266 2025

  43. [43]

    Ethics is not neutral

    Tiera Tanksley, Angela D. R. Smith, Saloni Sharma, and Earl W Huff. 2025. "Ethics is not neutral": Understanding Ethical and Responsible AI Design from the Lenses of Black Youth. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, 1–20. https://doi.org/10.1145/37...

    arXiv 2025

  44. [44]

    Joana Varon, Sasha Costanza-Chock, and Timnit Gebru. 2024. Fostering a Federated AI Commons ecosystem. (2024)

    2024

  45. [45]

    Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbhahn. 2024. Position: Will we run out of data? Limits of LLM scaling based on human-generated data. https://openreview.net/forum?id=Vi ZcgDQjyG

    2024

  46. [46]

    Christian Wagner and Ling Jiang. 2025. Death by AI: Will large language models diminish Wikipedia?Journal of the Association for Information Science and Technology76, 5 (2025), 743–751. https://doi.org/10.1002/asi.24975 _eprint: https://asistdl.onlinelibrary.wiley.com/doi/pdf/10.1002/asi.24975

    work page doi:10.1002/asi.24975 2025

  47. [47]

    Tianyu Wang, Akira Horiguchi, Lingyou Pang, and Carey E. Priebe. 2025. LLM Web Dynamics: Tracing Model Collapse in a Network of LLMs. https://doi.org/10.48550/arXiv.2506.15690 arXiv:2506.15690 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.15690 2025

  48. [48]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al . 2019. Huggingface’s transformers: State-of-the-art natural language processing.arXiv preprint arXiv:1910.03771(2019)

    Pith/arXiv arXiv 2019

  49. [49]

    Dustin Wright, Sarah Masud, Jared Moore, Srishti Yadav, Maria Antoniak, Chan Young Park, and Isabelle Augenstein

  50. [50]

    Jingyi Wu and Cailin O’Connor. 2023. How should we promote transient diversity in science?Synthese201, 2 (Jan. 2023). https://doi.org/10.1007/s11229-023-04037-1 Publisher: Springer Science and Business Media LLC

    work page doi:10.1007/s11229-023-04037-1 2023

  51. [51]

    Sang Michael Xie, Shibani Santurkar, Tengyu Ma, and Percy S. Liang. 2023. Data Selection for Language Models via Importance Resampling.Advances in Neural Information Processing Systems36 (Dec. 2023), 34201–34227. https://proc eedings.neurips.cc/paper_files/paper/2023/hash/6b9aa8f418bde2840d5f4ab7a02f663b-Abstract-Conference.html

    2023

  52. [52]

    Xiaodan Xing, Fadong Shi, Jiahao Huang, Yinzhe Wu, Yang Nan, Sheng Zhang, Yingying Fang, Michael Roberts, Carola-Bibiane Schönlieb, Javier Del Ser, and Guang Yang. 2025. On the caveats of AI autophagy.Nature Machine Intelligence7, 2 (Feb. 2025), 172–180. https://doi.org/10.1038/s42256-025-00984-1 Publisher: Nature Publishing Group

    doi:10.1038/s42256-025-00984-1 2025

  53. [53]

    Lily Hong Zhang, Smitha Milli, Karen Jusko, Jonathan Smith, Brandon Amos, Wassim Bouaziz, Manon Revel, Jack Kussman, Yasha Sheynin, Lisa Titus, Bhaktipriya Radharapu, Jane Yu, Vidya Sarma, Kris Rose, and Maximilian Nickel. 2025. Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset. https: //doi.org/10.48550/arXiv.2507.09650 ar...

    doi:10.48550/arxiv.2507.09650 2025

  54. [54]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022. OPT: Open Pre-trained Transformer Language Models. https://doi.org/10.48550...

    Pith/arXiv arXiv doi:10.48550/arxiv.2205.01068 2022

  55. [55]

    Xulu Zhang, Xiaoyong Wei, Jinlin Wu, Jiaxin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. 2025. Generating on Generated: An Approach Towards Self-Evolving Diffusion Models. https://doi.org/10.48550/arXiv.2502.09963 arXiv:2502.09963 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.09963 2025

  56. [56]

    Yuchang Zhu, Huazhen Zhong, Qunshu Lin, Haotong Wei, Xiaolong Sun, Zixuan Yu, Minghao Liu, Zibin Zheng, and Liang Chen. 2025. What Matters in LLM-generated Data: Diversity and Its Effect on Model Fine-Tuning. https: //doi.org/10.48550/arXiv.2506.19262 arXiv:2506.19262 [cs]. 14 Damian Hodel and Jevin D. West

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.19262 2025

  57. [57]

    Kevin J. S. Zollman. 2010. The Epistemic Benefit of Transient Diversity.Erkenntnis72, 1 (Jan. 2010), 17–35. https: //doi.org/10.1007/s10670-009-9194-6 Epistemic diversity across language models mitigates knowledge collapse 15 A Probability distributions Figure 6 illustrates the probability distributions of the information space. Information Information In...

    doi:10.1007/s10670-009-9194-6 2010

  58. [2024]

    https://doi.org/10.48550/arXiv.2404.01413 arXiv:2404.01413 [cs]

    Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data. https://doi.org/10.48550/arXiv.2404.01413 arXiv:2404.01413 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.01413

  59. [2025]

    https://doi.org/10.48550/arXiv.2510.04 226 arXiv:2510.04226 [cs]

    Epistemic Diversity and Knowledge Collapse in Large Language Models. https://doi.org/10.48550/arXiv.2510.04 226 arXiv:2510.04226 [cs]

    work page doi:10.48550/arxiv.2510.04