Growing a Tail: Increasing Output Diversity in Large Language Models
Open accessPublished version of a preprint
Abstract
How diverse are the outputs of large language models when diversity is desired? We examine the diversity of responses of various models to questions with multiple possible answers, comparing them with human responses. Our findings suggest that models' outputs are highly concentrated, reflecting a narrow, mainstream 'worldview', in comparison to humans, whose responses exhibit a much longer-tail. We examine three ways to increase models' output diversity: 1) increasing generation randomness via temperature sampling; 2) prompting models to answer from diverse perspectives; 3) aggregating outputs from several models. A combination of these measures significantly increases models' output diversity, reaching that of humans. We discuss implications of these findings for AI policy that wishes to preserve cultural diversity, an essential building block of a democratic social fabric.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Some work studies output variability via temperature scaling or decoding methods, but rarely distinguishes between stylistic variation, stochasticity, and knowledge diversity.
“Some work studies output variability via temperature scaling (Shur-Ofry et al., 2024;Peeperkorn et al., 2024) or decoding methods (Park et al., 2025;Su et al., 2022), but rarely distinguishes between stylistic variation, stochasticity, and knowledge diversity.”
From On Epistemic Diversity in Large Language Models (2026) · cites this paper as Shur-Ofry et al., 2024
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
A Mathematical Theory of Communication
Shannon, C. E. (1948) · Bell System Technical Journal · doi:10.1002/j.1538-7305.1948.tb01338.x · Cited by 88,486 · see citing works
LeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126 · see citing works
Chain-Of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1800 · Cited by 4,505 · see citing works
Benhabib, S. (2002) · Princeton University Press eBooks · doi:10.1515/9780691186542 · Cited by 1,894 · see citing works
Toolformer: language models can teach themselves to use tools
Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023) · Advances in Neural Information Processing Systems 36 · doi:10.52202/075280-2997 · Cited by 1,150 · see citing works
Gemini: A Family of Highly Capable Multimodal Models
Gemini Team, Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. (2023) · arXiv · doi:10.48550/arxiv.2312.11805 · Cited by 990 · see citing works
TruthfulQA: Measuring How Models Mimic Human Falsehoods
Lin, S., Hilton, J., & Evans, O. (2022) · Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2022.acl-long.229 · Cited by 904 · see citing works
A Tutorial on Energy-Based Learning
LeCun, Y., Chopra, S., Hadsell, R. T., Ranzato, A., & Huang, F. J. (2006) · Cited by 812 · see citing works
Algorithmic monoculture and social welfare
Kleinberg, J., & Raghavan, M. (2021) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2018340118 · Cited by 147 · see citing works
Language Models as Agent Models
Andreas, J. (2022) · Findings of the Association for Computational Linguistics: EMNLP 2022 · doi:10.18653/v1/2022.findings-emnlp.423 · Cited by 92 · see citing works
Picking on the Same Person: Does Algorithmic Monoculture Lead to Outcome Homogenization?
Bommasani, R., Creel, K., Kumar, A., Jurafsky, D., & Liang, P. (2022) · neural information processing systems · doi:10.52202/068431-0265 · Cited by 74 · see citing works
COLD Decoding: Energy-Based Constrained Text Generation with Langevin Dynamics
Qin, L., Welleck, S., Khashabi, D., & Choi, Y. (2022) · neural information processing systems · doi:10.52202/068431-0693 · Cited by 53 · see citing works
Trading Off Diversity and Quality in Natural Language GenerationPreprint
Zhang, H., Duckworth, D., Ippolito, D., & Neelakantan, A. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2004.10450 · Cited by 50 · see citing works
On Decoding Strategies for Neural Text Generators
Wiher, G., Meister, C., & Cotterell, R. (2022) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00502 · Cited by 48 · see citing works
Towards Measuring the Representation of Subjective Global Opinions in Language Models
Durmus, E., Nguyen, K., Liao, T. I., Schiefer, N., Askell, A., Bakhtin, A., Chen, C., Hatfield-Dodds, Z., Hernandez, D., Joseph, N., et al. (2023) · arXiv · doi:10.48550/arxiv.2306.16388 · Cited by 48 · see citing works
Controlled Text Generation as Continuous Optimization with Multiple ConstraintsPreprint
Kumar, S., Malmi, E., Severyn, A., & Tsvetkov, Y. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2108.01850 · Cited by 29 · see citing works
Gradient-based Constrained Sampling from Language Models
Kumar, S., Paria, B., & Tsvetkov, Y. (2022) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2022.emnlp-main.144 · Cited by 28 · see citing works
Democratic Constitutionalism and Cultural HeterogeneityPreprint
Post, R. C. (2000) · SSRN Electronic Journal · doi:10.2139/ssrn.229161 · Cited by 19 · see citing works
A Systematic Characterization of Sampling Algorithms for Open-ended Language GenerationPreprint
Nadeem, M., He, T., Cho, K., & Glass, J. R. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2009.07243 · Cited by 8 · see citing works
Multiplicity as an AI Governance PrinciplePreprint
Shur‐Ofry, M. (2023) · SSRN Electronic Journal · doi:10.2139/ssrn.4444354 · Cited by 6 · see citing works
Self-citationDiverse Preference Optimization
Lanchantin, J., Chen, A., Dhuliawala, S., Yu, P., Weston, J., Sukhbaatar, S., & Kulikov, I. (2025) · arXiv · doi:10.48550/arxiv.2501.18101 · Cited by 2 · see citing works
Modifying Large Language Model Post-Training for Diverse Creative WritingPreprint
Chung, J. J. Y., Padmakumar, V., Roemmele, M., Sun, Y., & Kreminski, M. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2503.17126 · Cited by 1 · see citing works
Turning up the heat: Min-p sampling for creative and coherent llm outputs
Nguyen, M. N., Baker, A., Neo, C., Roush, A., Kirsch, A., & Shwartz-Ziv, R. (2024) · arXiv · doi:10.48550/arxiv.2407.01082 · Cited by 1 · see citing works
Personas as a Way to Model Truthfulness in Language ModelsPreprint
Joshi, N., Rando, J., Saparov, A., Kim, N., & He, H. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2310.18168
Lagzian, A., Anumasa, S., & Liu, D. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2502.12700
Forcing diffuse distributions out of language models
Zhang, Y., Schwarzschild, A., Carlini, N., Kolter, Z., & Ippolito, D. (2024) · arXiv · doi:10.48550/arxiv.2404.10859
Tool Learning with Large Language Models: A Survey
Qu, C., Dai, S., Wei, X., et al. (2024) · arXiv · doi:10.48550/arxiv.2405.17935
Flagged first, then by citations
Research works printed but not linked: 13
- Achiam, O. J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., … Zoph, B. (2023). GPT-4 technical report. arXiv. https://arxiv.org/abs/2303.08774s.
- Baker, C. E. (2002). Media, markets, and democracy. Cambridge University Press.
- Corse, S. M. (1997). Nationalism and literature: The politics of culture in Canada and the United States. Cambridge University Press.
- Franceschelli, G., & Musolesi, M. (2025). DiffSampling: Enhancing diversity and accuracy in neural text generation. arXiv https://doi.org/10.48550/arXiv.2502.14037. · DOI
- Hayati, S. A., Lee., M., & Rajagopal, D. & Kang, D. (2024). How far can we extract diverse perspectives from large language models? Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 5336-5366). Association for Computational Linguistics.
- Li, Z., Chen, C., Xu, T., Qin, Z., Xiao, J., Luo, Z. & Sun, R. (2025). Preserving diversity in supervised fine-tuning of large language models. In International Conference on Learning Representations. arXiv. https://doi.org/10.48550/arXiv.2408.16673. · DOI
- Mai, L., & Carson-Berndsen, J. (2024). Improving linguistic diversity of large language models with possibility exploration fine-tuning. arXiv. https://doi.org/10.48550/arXiv.2412.03343. · DOI
- Napoli, P. M. (2011). Exposure diversity reconsidered. Journal of Information Policy, 1, 246-259.
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence. (2024). Official Journal of the European Union, L series,2024/1689,1. https://eur-lex.europa.eu/legal- ontent/EN/TXT/?uri=OJ:L_202401689.
- Tekin, S. F., Ilhan, F., Huang, T., Hu, S.& Liu, L. (2024). LLM-TOPLA: Efficient LLM ensemble by maximising diversity. Findings of the Association for Computational Linguistics: EMNLP 2024 (pp. 11951- 11966). Association for Computational Linguistics.
- Wong, J., Orlovsky, Y., Luo, M., Seshia, S., & Gonzalez, J.E. (2024). SimpleStrat: Diversifying language model generation with stratification. arXiv. https://doi.org/10.48550/arXiv.2410.09038. · DOI
- Yadav, V., Kwon, H., Srinivasan, V., & Jin, H. (2024). Explicit diversity conditions for effective question answer generation with large language models. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 6876-6882).
- Yu, Y., Zhuang, Y., Zhang, J., Meng, Y., Ratner, A., Krishna, R., Shen, J., & Zhang, C. (2023). Large language model as attributed training data generator: A tale of diversity and bias. Advances in Neural Information Processing Systems, 36, 55734-55784.
Cited by
Sorted by citations
ChatGPT is incredible (at being average)
Rudko, I., & Bonab, A. B. (2025) · Ethics and Information Technology · doi:10.1007/s10676-025-09845-2 · Cited by 26 · see citing works
Fair or flawed? Rethinking grant review with generative AI
Sikimić, V. (2025) · Synthese · doi:10.1007/s11229-025-05366-z · Cited by 2 · see citing works
On Epistemic Diversity in Large Language Models
Kirsten, E., Krämer, N., & Zafar, M. B. (2026) · arXiv · doi:10.48550/arxiv.2609.04835
The (non-)alienness-dilemma in the use of simulated research subjects in psychological research
Abel, L. (2026) · Philosophical Psychology · doi:10.1080/09515089.2026.2752237
- Authors
- Shur-Ofry, M., Horowitz-Amsalem, B., Rahamim, A., & Belinkov, Y.
- Affiliations
- Hebrew University of Jerusalem (IL)
- Technion – Israel Institute of Technology (IL)
- Corresponding author
- Belinkov, Yonatan
- ORCID
- Belinkov, Yonatan
- Venue
- Machine Learning with Applications
- Publisher
- Elsevier BV
- Published
- 2026-09
- Volume
- 25
- Pages
- 100951
- ISSN
- 2666-8270 (ISSN-L)
- OpenAlex
- W4404395419
- Cited by
- 3
- References
- 18
- Citations by year
- 2026: 1 · 2025: 2
- FWCI
- 3.42
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Machine Learning with Applications (publishedVersion, cc-by-nc-nd)
- SSRN Electronic Journal (acceptedVersion)
- arXiv (Cornell University) (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- arXiv (Cornell University)
- Funders
- Israel Science Foundation
- European Research Council
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Social Sciences › Safety Research › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- output diversity, large language models, temperature sampling, output aggregation, cultural diversity, AI governance
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
