Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities

Benjamin Rosman; Devon Jarvis; Richard Klein; Stefano Sarao Mannelli; Steven James

REVIEW 2 major objections 1 minor 34 references

Model collapse threatens AI democratization by skewing data away from rare patterns that low-resource communities rely on.

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 · grok-4.3

2026-07-01 00:07 UTC pith:RY2OCFXP

load-bearing objection Position paper links model collapse to equity harms but rests on an untested assumption that collapse erodes tail mass more for low-resource data. the 2 major comments →

arxiv 2605.04127 v2 pith:RY2OCFXP submitted 2026-05-05 cs.LG cs.CLcs.CY

Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities

Devon Jarvis , Richard Klein , Benjamin Rosman , Steven James , Stefano Sarao Mannelli This is my paper

classification cs.LG cs.CLcs.CY

keywords model collapseAI democratizationlow-resource communitiessynthetic datadata degradationcultural biasenvironmental cost

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 argues that model collapse, the performance drop when models train on outputs from earlier models, will reduce training efficiency and push data distributions away from their less common tails. This effect hits low-resource and marginalized groups hardest because those communities depend more on the full range of patterns rather than the most frequent ones. The authors link the issue to both environmental waste from inefficient training and the reinforcement of cultural biases, and they call for targeted mitigation to protect access to useful AI.

Core claim

Model collapse threatens current efforts to democratize AI. By reducing training efficiency and skewing data distributions away from the tails of their support, model collapse disproportionately impacts low-resource and marginalized communities.

What carries the argument

Model collapse, defined as performance degradation that occurs when generative models are trained on the synthetic outputs of prior models.

Load-bearing premise

Model collapse will systematically reduce representation of the less common patterns in data in ways that affect low-resource communities more than others.

What would settle it

Empirical measurement of tail-support coverage in training corpora for low-resource languages before and after large-scale introduction of model-generated text, showing no net loss or a gain for those languages.

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

If this is right

Training runs on synthetic data will require more compute to reach the same performance, raising costs for everyone but especially for groups with fewer resources.
Rare cultural and linguistic patterns will receive less weight in future models, narrowing the effective coverage for communities outside the data majority.
Environmental costs of AI development will rise without delivering proportional gains in usefulness for low-resource settings.
Cultural biases already present in data will be amplified as the model outputs increasingly favor high-frequency patterns.

Where Pith is reading between the lines

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

Preserving large stores of original human-generated data may become a practical requirement for maintaining broad model utility.
Similar tail-erasure effects could appear in non-text domains such as image or code generation when synthetic data dominates.
Monitoring the diversity of publicly available training corpora over time could serve as an early warning system for the predicted skew.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Position paper links model collapse to equity harms but rests on an untested assumption that collapse erodes tail mass more for low-resource data.

read the letter

The main thing here is a framing argument: model collapse, already flagged in prior work, will hit low-resource communities harder because synthetic data training supposedly strips out the tails of distributions that matter most to them. The paper pulls together existing points on collapse, bias amplification, and resource costs, then adds the equity angle and a call to action.

What it does cleanly is connect the dots between three strands that usually sit in separate papers—collapse mechanics, cultural skew, and environmental load—and points out that repeated synthetic-data loops could make it even harder for groups with small original datasets to catch up. That synthesis is straightforward and the environmental/cultural implications section follows logically from the cited literature.

The soft spot is the load-bearing claim that collapse systematically moves mass away from the tails and that this effect is larger for marginalized data. The abstract states this without a model, simulation, or even a sketch of why tail erosion would be differential rather than uniform. No comparative numbers across resource levels appear, and the stress-test note is right that the causal step is assumed rather than derived. As a position paper this is acceptable for raising the hypothesis, but it leaves the central mechanism unexamined.

This is for readers already working on AI ethics, data governance, or low-resource NLP who want a compact way to link collapse to fairness arguments. It is not carrying new technical results, so it will not change how most labs think about training pipelines.

I would bring it to a reading group if the group is discussing equity or synthetic data risks, but I would not cite it for any technical claim. It deserves peer review in an ethics or policy venue to get the hypothesis on record and invite the missing empirical work, even though the current version is light on evidence.

Referee Report

2 major / 1 minor

Summary. This position paper argues that model collapse—the performance degradation from training generative models on synthetic outputs—threatens efforts to democratize AI. It claims that reduced training efficiency combined with systematic skewing of data distributions away from the tails of their support disproportionately harms low-resource and marginalized communities, while also examining environmental and cultural implications and issuing a call to action for mitigation strategies.

Significance. If the core causal links hold, the synthesis could usefully connect technical concerns about synthetic data with equity issues in AI access, potentially informing data governance and training practices for underrepresented groups. The paper draws on existing literature on model collapse and bias but offers no new derivations, empirical tests, or reproducible elements to strengthen the argument.

major comments (2)

[Abstract] Abstract: the load-bearing claim that model collapse 'skewing data distributions away from the tails of their support' produces a directional effect that 'disproportionately impacts low-resource and marginalized communities' is asserted without a formal model, cited mechanism, or comparative evidence showing tail erosion is larger or more consequential for low-resource data regimes than for high-resource ones.
[Abstract] Abstract / main argument: the linkage between reduced training efficiency, tail-mass loss, and differential community impact is presented as following directly from cited critiques of LLMs, yet no derivation or reference establishes why synthetic-data training would systematically erode tails rather than produce uniform degradation or mode collapse.

minor comments (1)

The manuscript would benefit from explicit separation of claims drawn from prior work versus the novel synthesis regarding low-resource communities.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our position paper. Below we respond point-by-point to the major comments, maintaining the scope of a position paper that synthesizes existing literature rather than introducing new formal models or empirical results.

read point-by-point responses

Referee: [Abstract] Abstract: the load-bearing claim that model collapse 'skewing data distributions away from the tails of their support' produces a directional effect that 'disproportionately impacts low-resource and marginalized communities' is asserted without a formal model, cited mechanism, or comparative evidence showing tail erosion is larger or more consequential for low-resource data regimes than for high-resource ones.

Authors: As a position paper, the manuscript synthesizes arguments from the model collapse literature and existing LLM critiques to posit this directional effect, rather than deriving a new formal model or presenting comparative evidence. The skewing claim is motivated by documented tendencies of synthetic training data to reinforce frequent patterns at the expense of diversity. We acknowledge that no new mechanism or low- versus high-resource comparison is supplied. We will revise the abstract to frame the claim more explicitly as a synthesized hypothesis warranting further study. revision: partial
Referee: [Abstract] Abstract / main argument: the linkage between reduced training efficiency, tail-mass loss, and differential community impact is presented as following directly from cited critiques of LLMs, yet no derivation or reference establishes why synthetic-data training would systematically erode tails rather than produce uniform degradation or mode collapse.

Authors: The linkage rests on the cited critiques, which emphasize reproduction of frequent patterns, and on model collapse studies that report reduced output diversity. These sources indicate effects that go beyond uniform degradation. No new derivation is offered, consistent with the position-paper format. We can expand the references in a revision to include additional works that specifically address tail erosion under synthetic data regimes. revision: partial

Circularity Check

0 steps flagged

No circularity: position paper advances conceptual argument from external citations without derivations or self-referential reductions

full rationale

The manuscript is a position paper containing no equations, parameters, or formal derivations. Its central claim—that model collapse skews distributions away from tails and disproportionately affects low-resource communities—is presented as a hypothesis combining views from cited literature on model collapse, data degradation, and cultural biases. No step reduces a prediction or result to a fitted input, self-citation chain, or definitional equivalence by construction. Self-citations (if present) are not load-bearing for any mathematical claim, and the argument remains self-contained against external benchmarks without internal circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that model collapse produces specific effects on efficiency and tail distributions that map directly to disproportionate harm for low-resource groups; no free parameters or invented entities are introduced.

axioms (1)

domain assumption Model collapse reduces training efficiency and skews data distributions away from the tails of their support.
This premise is invoked to link collapse to disproportionate impacts on low-resource communities.

pith-pipeline@v0.9.1-grok · 5697 in / 1028 out tokens · 37601 ms · 2026-07-01T00:07:47.452126+00:00 · methodology

0 comments

read the original abstract

Model collapse, the degradation in performance that arises when generative models are trained on the outputs of prior models, is an increasing concern as artificially generated content proliferates. Related critiques of large language models have highlighted their tendency to reproduce frequent patterns in training data, their reliance on vast datasets, and their substantial environmental cost. Together, these factors contribute to data degradation, the reinforcement of cultural biases, and inefficient resource use. In this position paper we aim to combine these views and argue that model collapse threatens current efforts to democratize AI. By reducing training efficiency and skewing data distributions away from the tails of their support, model collapse disproportionately impacts low-resource and marginalized communities. We examine both the environmental and cultural implications of this phenomenon, situate our position within recent position papers on model collapse, and conclude with a call to action. Finally, we outline initial directions for mitigating these effects.

Figures

Figures reproduced from arXiv: 2605.04127 by Benjamin Rosman, Devon Jarvis, Richard Klein, Stefano Sarao Mannelli, Steven James.

**Figure 1.** Figure 1: Perplexity of multiple languages using the Latin alphabet (potentially with some added characters) calculated using a pretrained GPT-2 (Radford et al., 2019). Note how the lower-resource languages occupy a distribution closer to the tails (at a higher perplexity) than the more high-resource languages such as English. Each language distribution is calculated using 20000 input sentences (agentlans, 2025) … view at source ↗

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 9 canonical work pages · 3 internal anchors

[1]

M., Gebru, T., McMillan-Major, A., and Shmitchell, S

Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? InProceedings of the 2021 ACM conference on fairness, accountability, and trans- parency, pp. 610–623,

2021
[2]

and Farid, H

Bohacek, M. and Farid, H. Nepotistically trained generative image models collapse. InICLR 2025 Workshop on Nav- igating and Addressing Data Problems for Foundation Models,

2025
[3]

S., Singh, J., and Anastasopoulos, A

Choi, A., Akter, S. S., Singh, J., and Anastasopoulos, A. The llm effect: Are humans truly using llms, or are they being influenced by them instead? InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 22032–22054,

2024
[4]

Reducing Overfitting in Deep Networks by Decorrelating Representations

Cogswell, M., Ahmed, F., Girshick, R., Zitnick, L., and Ba- tra, D. Reducing overfitting in deep networks by decorre- lating representations.arXiv preprint arXiv:1511.06068,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Bert: Pre-training of deep bidirectional transformers for lan- guage understanding

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for lan- guage understanding. InProceedings of the 2019 confer- ence of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186,

2019
[6]

Model collapse demystified: The case of regression.Advances in Neu- ral Information Processing Systems, 37:46979–47013, 2024a

Dohmatob, E., Feng, Y ., and Kempe, J. Model collapse demystified: The case of regression.Advances in Neu- ral Information Processing Systems, 37:46979–47013, 2024a. Dohmatob, E., Feng, Y ., Subramonian, A., and Kempe, J. Strong model collapse.arXiv preprint arXiv:2410.04840, 2024b. Dohmatob, E., Feng, Y ., Yang, P., Charton, F., and Kempe, J. A tale of t...

work page arXiv
[7]

Analysis and forecast to 2026.International Energy Agency: Paris, France,

Electricity, I. Analysis and forecast to 2026.International Energy Agency: Paris, France,

2026
[8]

Po- sition: Cracking the code of cascading disparity to- wards marginalized communities

Farnadi, G., Havaei, M., and Rostamzadeh, N. Po- sition: Cracking the code of cascading disparity to- wards marginalized communities. InForty-first Inter- national Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27,

2024
[9]

Feng, Y ., Dohmatob, E., Yang, P., Charton, F., and Kempe, J

URL https://openreview.net/forum? id=XDz9leJ9iK. Feng, Y ., Dohmatob, E., Yang, P., Charton, F., and Kempe, J. Beyond model collapse: Scaling up with synthesized data requires reinforcement. InICML 2024 workshop on theoretical foundations of foundation models,

2024
[10]

Do llms exhibit human-like cognitive biases? a large-scale sys- tematic evaluation.A Large-Scale Systematic Evaluation (September 17, 2025),

Geva, T., Goldstein, A., Lary, E., and Levy, C. Do llms exhibit human-like cognitive biases? a large-scale sys- tematic evaluation.A Large-Scale Systematic Evaluation (September 17, 2025),

2025
[11]

Social bias evalua- tion for large language models requires prompt variations

Hida, R., Kaneko, M., and Okazaki, N. Social bias evalua- tion for large language models requires prompt variations. InFindings of the Association for Computational Linguis- tics: EMNLP 2025, pp. 14507–14530,

2025
[12]

Jegham, S

Jegham, N., Abdelatti, M., Koh, C. Y ., Elmoubarki, L., and Hendawi, A. How hungry is ai? benchmarking en- ergy, water, and carbon footprint of llm inference.arXiv preprint arXiv:2505.09598,

work page arXiv
[13]

Jo, E. S. and Gebru, T. Lessons from archives: Strategies for collecting sociocultural data in machine learning. In Proceedings of the 2020 conference on fairness, account- ability, and transparency, pp. 306–316,

2020
[14]

Scaling Laws for Neural Language Models

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361,

work page internal anchor Pith review Pith/arXiv arXiv 2001
[15]

Kirsten, E., Habernal, I., Nanda, V ., and Zafar, M. B. The impact of inference acceleration on bias of llms. InPro- ceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 1834–1853,

2025
[16]

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task

Kosmyna, N., Hauptmann, E., Yuan, Y . T., Situ, J., Liao, X.-H., Beresnitzky, A. V ., Braunstein, I., and Maes, P. Your brain on chatgpt: Accumulation of cognitive debt when using an ai assistant for essay writing task.arXiv preprint arXiv:2506.08872, 4,

work page internal anchor Pith review Pith/arXiv arXiv
[17]

and Richardson, J

Kudo, T. and Richardson, J. Sentencepiece: A simple and language independent subword tokenizer and detok- enizer for neural text processing. InProceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations. Association for Computational Linguistics,

2018
[18]

D., Ngo, N., Veyseh, A

Lai, V . D., Ngo, N., Veyseh, A. P. B., Man, H., Dernoncourt, F., Bui, T., and Nguyen, T. H. Chatgpt beyond english: Towards a comprehensive evaluation of large language models in multilingual learning. InFindings of the asso- ciation for computational linguistics: EMNLP 2023, pp. 13171–13189,

2023
[19]

Countering language drift via visual grounding

Lee, J., Cho, K., and Kiela, D. Countering language drift via visual grounding. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 4385–4395,

2019
[20]

Regmix: Data mixture as regression for language model pre-training

Liu, Q., Zheng, X., Muennighoff, N., Zeng, G., Dou, L., Pang, T., Jiang, J., and Lin, M. Regmix: Data mixture as regression for language model pre-training. InInterna- tional Conference on Learning Representations, volume 2025, pp. 38305–38339,

2025
[21]

arXiv preprint arXiv:2506.14012 , year=

Mohamed, A., Zhang, Y ., Vazirgiannis, M., and Shang, G. Lost in the mix: Evaluating llm understanding of code- switched text.arXiv preprint arXiv:2506.14012,

work page arXiv
[22]

com/2025/05/20/1116327/ ai-energy-usage-climate-footprint-big-tech/

URL https://www.technologyreview. com/2025/05/20/1116327/ ai-energy-usage-climate-footprint-big-tech/ . Olaleye, K., Oncevay, A., Sibue, M., Zondi, N., Terblanche, M., Mapikitla, S., Lastrucci, R., Smiley, C., and Marivate, 12 Position: the Stochastic Parrot in the Coal Mine V . Afrocs-xs: Creating a compact, high-quality, human- validated code-switched d...

2025
[23]

Llms know more than they show: On the intrinsic representation of llm hallucinations

Orgad, H., Toker, M., Gekhman, Z., Reichart, R., Szpek- tor, I., Kotek, H., and Belinkov, Y . Llms know more than they show: On the intrinsic representation of llm hallucinations. InInternational Conference on Learning Representations, volume 2025, pp. 66880–66913,

2025
[24]

S., Khandelwal, A., Tanmay, K., Agarwal, U., and Choudhury, M

Rao, A. S., Khandelwal, A., Tanmay, K., Agarwal, U., and Choudhury, M. Ethical reasoning over moral alignment: A case and framework for in-context ethical policies in llms. InFindings of the Association for Computational Linguistics: EMNLP 2023, pp. 13370–13388,

2023
[25]

Perturbation bounds for means of eigenvalues and invariant subspaces.BIT Numerical Mathematics, 10(3): 343–354, 1970

Schaeffer, R., Kazdan, J., Arulandu, A. C., and Koyejo, S. Position: Model collapse does not mean what you think. arXiv preprint arXiv:2503.03150,

work page arXiv
[26]

what shapes your bias?

Shin, J., Song, H., Lee, H., Jeong, S., and Park, J. C. Ask llms directly,“what shapes your bias?”: Measuring so- cial bias in large language models. InFindings of the Association for Computational Linguistics ACL 2024, pp. 16122–16143,

2024
[27]

Towards a comprehensive understanding and accurate evaluation of societal biases in pre-trained transformers

Silva, A., Tambwekar, P., and Gombolay, M. Towards a comprehensive understanding and accurate evaluation of societal biases in pre-trained transformers. InPro- ceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 2383–2389,

2021
[28]

Spennemann, D. H. Delving into: the quantification of ai- generated content on the internet (synthetic data).arXiv preprint arXiv:2504.08755,

work page arXiv
[29]

Stankovic, M., Hirche, E., Kollatzsch, S., and Doetsch, J. N. Comment on: Your brain on chatgpt: Accumulation of cognitive debt when using an ai assistant for essay writing tasks.arXiv preprint arXiv:2601.00856,

work page arXiv
[30]

Thudi, A., Rovers, E., Ruan, Y ., Thrush, T., and Maddison, C. J. Mixmin: Finding data mixtures via convex mini- mization. InForty-second International Conference on Machine Learning. Twyman, M., Keegan, B. C., and Shaw, A. Black lives matter in wikipedia: Collective memory and collaboration around online social movements. InProceedings of the 2017 acm co...

2017
[31]

Whitney, C. D. and Norman, J. Real risks of fake data: Syn- thetic data, diversity-washing and consent circumvention. InProceedings of the 2024 ACM conference on fairness, accountability, and transparency, pp. 1733–1744,

2024
[32]

I., Aji, A

Winata, G. I., Aji, A. F., Yong, Z.-X., and Solorio, T. The decades progress on code-switching research in nlp: A systematic survey on trends and challenges.Findings of the Association for Computational Linguistics: ACL 2023, pp. 2936–2978,

2023
[33]

Fairness feed- back loops: training on synthetic data amplifies bias

Wyllie, S., Shumailov, I., and Papernot, N. Fairness feed- back loops: training on synthetic data amplifies bias. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 2113–2147,

2024
[34]

Code- switching curriculum learning for multilingual transfer in llms

Yoo, H., Park, C., Yun, S., Oh, A., and Lee, H. Code- switching curriculum learning for multilingual transfer in llms. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 7816–7836, 2025a. Yoo, H., Yang, Y ., and Lee, H. Code-switching red-teaming: Llm evaluation for safety and multilingual understanding. InProceedings of the 63rd A...

2025

[1] [1]

M., Gebru, T., McMillan-Major, A., and Shmitchell, S

Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? InProceedings of the 2021 ACM conference on fairness, accountability, and trans- parency, pp. 610–623,

2021

[2] [2]

and Farid, H

Bohacek, M. and Farid, H. Nepotistically trained generative image models collapse. InICLR 2025 Workshop on Nav- igating and Addressing Data Problems for Foundation Models,

2025

[3] [3]

S., Singh, J., and Anastasopoulos, A

Choi, A., Akter, S. S., Singh, J., and Anastasopoulos, A. The llm effect: Are humans truly using llms, or are they being influenced by them instead? InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 22032–22054,

2024

[4] [4]

Reducing Overfitting in Deep Networks by Decorrelating Representations

Cogswell, M., Ahmed, F., Girshick, R., Zitnick, L., and Ba- tra, D. Reducing overfitting in deep networks by decorre- lating representations.arXiv preprint arXiv:1511.06068,

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

Bert: Pre-training of deep bidirectional transformers for lan- guage understanding

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for lan- guage understanding. InProceedings of the 2019 confer- ence of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186,

2019

[6] [6]

Model collapse demystified: The case of regression.Advances in Neu- ral Information Processing Systems, 37:46979–47013, 2024a

Dohmatob, E., Feng, Y ., and Kempe, J. Model collapse demystified: The case of regression.Advances in Neu- ral Information Processing Systems, 37:46979–47013, 2024a. Dohmatob, E., Feng, Y ., Subramonian, A., and Kempe, J. Strong model collapse.arXiv preprint arXiv:2410.04840, 2024b. Dohmatob, E., Feng, Y ., Yang, P., Charton, F., and Kempe, J. A tale of t...

work page arXiv

[7] [7]

Analysis and forecast to 2026.International Energy Agency: Paris, France,

Electricity, I. Analysis and forecast to 2026.International Energy Agency: Paris, France,

2026

[8] [8]

Po- sition: Cracking the code of cascading disparity to- wards marginalized communities

Farnadi, G., Havaei, M., and Rostamzadeh, N. Po- sition: Cracking the code of cascading disparity to- wards marginalized communities. InForty-first Inter- national Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27,

2024

[9] [9]

Feng, Y ., Dohmatob, E., Yang, P., Charton, F., and Kempe, J

URL https://openreview.net/forum? id=XDz9leJ9iK. Feng, Y ., Dohmatob, E., Yang, P., Charton, F., and Kempe, J. Beyond model collapse: Scaling up with synthesized data requires reinforcement. InICML 2024 workshop on theoretical foundations of foundation models,

2024

[10] [10]

Do llms exhibit human-like cognitive biases? a large-scale sys- tematic evaluation.A Large-Scale Systematic Evaluation (September 17, 2025),

Geva, T., Goldstein, A., Lary, E., and Levy, C. Do llms exhibit human-like cognitive biases? a large-scale sys- tematic evaluation.A Large-Scale Systematic Evaluation (September 17, 2025),

2025

[11] [11]

Social bias evalua- tion for large language models requires prompt variations

Hida, R., Kaneko, M., and Okazaki, N. Social bias evalua- tion for large language models requires prompt variations. InFindings of the Association for Computational Linguis- tics: EMNLP 2025, pp. 14507–14530,

2025

[12] [12]

Jegham, S

Jegham, N., Abdelatti, M., Koh, C. Y ., Elmoubarki, L., and Hendawi, A. How hungry is ai? benchmarking en- ergy, water, and carbon footprint of llm inference.arXiv preprint arXiv:2505.09598,

work page arXiv

[13] [13]

Jo, E. S. and Gebru, T. Lessons from archives: Strategies for collecting sociocultural data in machine learning. In Proceedings of the 2020 conference on fairness, account- ability, and transparency, pp. 306–316,

2020

[14] [14]

Scaling Laws for Neural Language Models

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361,

work page internal anchor Pith review Pith/arXiv arXiv 2001

[15] [15]

Kirsten, E., Habernal, I., Nanda, V ., and Zafar, M. B. The impact of inference acceleration on bias of llms. InPro- ceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 1834–1853,

2025

[16] [16]

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task

Kosmyna, N., Hauptmann, E., Yuan, Y . T., Situ, J., Liao, X.-H., Beresnitzky, A. V ., Braunstein, I., and Maes, P. Your brain on chatgpt: Accumulation of cognitive debt when using an ai assistant for essay writing task.arXiv preprint arXiv:2506.08872, 4,

work page internal anchor Pith review Pith/arXiv arXiv

[17] [17]

and Richardson, J

Kudo, T. and Richardson, J. Sentencepiece: A simple and language independent subword tokenizer and detok- enizer for neural text processing. InProceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations. Association for Computational Linguistics,

2018

[18] [18]

D., Ngo, N., Veyseh, A

Lai, V . D., Ngo, N., Veyseh, A. P. B., Man, H., Dernoncourt, F., Bui, T., and Nguyen, T. H. Chatgpt beyond english: Towards a comprehensive evaluation of large language models in multilingual learning. InFindings of the asso- ciation for computational linguistics: EMNLP 2023, pp. 13171–13189,

2023

[19] [19]

Countering language drift via visual grounding

Lee, J., Cho, K., and Kiela, D. Countering language drift via visual grounding. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 4385–4395,

2019

[20] [20]

Regmix: Data mixture as regression for language model pre-training

Liu, Q., Zheng, X., Muennighoff, N., Zeng, G., Dou, L., Pang, T., Jiang, J., and Lin, M. Regmix: Data mixture as regression for language model pre-training. InInterna- tional Conference on Learning Representations, volume 2025, pp. 38305–38339,

2025

[21] [21]

arXiv preprint arXiv:2506.14012 , year=

Mohamed, A., Zhang, Y ., Vazirgiannis, M., and Shang, G. Lost in the mix: Evaluating llm understanding of code- switched text.arXiv preprint arXiv:2506.14012,

work page arXiv

[22] [22]

com/2025/05/20/1116327/ ai-energy-usage-climate-footprint-big-tech/

URL https://www.technologyreview. com/2025/05/20/1116327/ ai-energy-usage-climate-footprint-big-tech/ . Olaleye, K., Oncevay, A., Sibue, M., Zondi, N., Terblanche, M., Mapikitla, S., Lastrucci, R., Smiley, C., and Marivate, 12 Position: the Stochastic Parrot in the Coal Mine V . Afrocs-xs: Creating a compact, high-quality, human- validated code-switched d...

2025

[23] [23]

Llms know more than they show: On the intrinsic representation of llm hallucinations

Orgad, H., Toker, M., Gekhman, Z., Reichart, R., Szpek- tor, I., Kotek, H., and Belinkov, Y . Llms know more than they show: On the intrinsic representation of llm hallucinations. InInternational Conference on Learning Representations, volume 2025, pp. 66880–66913,

2025

[24] [24]

S., Khandelwal, A., Tanmay, K., Agarwal, U., and Choudhury, M

Rao, A. S., Khandelwal, A., Tanmay, K., Agarwal, U., and Choudhury, M. Ethical reasoning over moral alignment: A case and framework for in-context ethical policies in llms. InFindings of the Association for Computational Linguistics: EMNLP 2023, pp. 13370–13388,

2023

[25] [25]

Perturbation bounds for means of eigenvalues and invariant subspaces.BIT Numerical Mathematics, 10(3): 343–354, 1970

Schaeffer, R., Kazdan, J., Arulandu, A. C., and Koyejo, S. Position: Model collapse does not mean what you think. arXiv preprint arXiv:2503.03150,

work page arXiv

[26] [26]

what shapes your bias?

Shin, J., Song, H., Lee, H., Jeong, S., and Park, J. C. Ask llms directly,“what shapes your bias?”: Measuring so- cial bias in large language models. InFindings of the Association for Computational Linguistics ACL 2024, pp. 16122–16143,

2024

[27] [27]

Towards a comprehensive understanding and accurate evaluation of societal biases in pre-trained transformers

Silva, A., Tambwekar, P., and Gombolay, M. Towards a comprehensive understanding and accurate evaluation of societal biases in pre-trained transformers. InPro- ceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 2383–2389,

2021

[28] [28]

Spennemann, D. H. Delving into: the quantification of ai- generated content on the internet (synthetic data).arXiv preprint arXiv:2504.08755,

work page arXiv

[29] [29]

Stankovic, M., Hirche, E., Kollatzsch, S., and Doetsch, J. N. Comment on: Your brain on chatgpt: Accumulation of cognitive debt when using an ai assistant for essay writing tasks.arXiv preprint arXiv:2601.00856,

work page arXiv

[30] [30]

Thudi, A., Rovers, E., Ruan, Y ., Thrush, T., and Maddison, C. J. Mixmin: Finding data mixtures via convex mini- mization. InForty-second International Conference on Machine Learning. Twyman, M., Keegan, B. C., and Shaw, A. Black lives matter in wikipedia: Collective memory and collaboration around online social movements. InProceedings of the 2017 acm co...

2017

[31] [31]

Whitney, C. D. and Norman, J. Real risks of fake data: Syn- thetic data, diversity-washing and consent circumvention. InProceedings of the 2024 ACM conference on fairness, accountability, and transparency, pp. 1733–1744,

2024

[32] [32]

I., Aji, A

Winata, G. I., Aji, A. F., Yong, Z.-X., and Solorio, T. The decades progress on code-switching research in nlp: A systematic survey on trends and challenges.Findings of the Association for Computational Linguistics: ACL 2023, pp. 2936–2978,

2023

[33] [33]

Fairness feed- back loops: training on synthetic data amplifies bias

Wyllie, S., Shumailov, I., and Papernot, N. Fairness feed- back loops: training on synthetic data amplifies bias. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 2113–2147,

2024

[34] [34]

Code- switching curriculum learning for multilingual transfer in llms

Yoo, H., Park, C., Yun, S., Oh, A., and Lee, H. Code- switching curriculum learning for multilingual transfer in llms. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 7816–7836, 2025a. Yoo, H., Yang, Y ., and Lee, H. Code-switching red-teaming: Llm evaluation for safety and multilingual understanding. InProceedings of the 63rd A...

2025