Elamur Research⌕
☰
home / papers / 2026 / CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters

CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters

Paper · 2026 · Ao Sun, Xiaoyu Wang, Zhe Tan et al.

Summary

Standard LLM alignment fits one shared set of weights to pooled human feedback. That works when a single correct answer exists, but it breaks down on culturally contested questions. The authors argue that the value distributions of different demographic groups form well-separated clusters, which they call Cultural Sparsity. Dense models trained on such data drift toward a bland averaged answer that often reflects WEIRD norms. They name this failure Mean Collapse and explain it theoretically through the mean-matching behaviour of forward-KL training.

Their remedy is CuMA (Cultural Mixture of Adapters). It inserts a mixture of adapter experts into the transformer FFN. A router picks experts from both the hidden states and an embedding of the user's demographic profile, which is rewritten as a short text and encoded by a frozen sentence-embedding model. Conflicting gradients therefore end up in separate expert subspaces.

On WorldValuesBench, Community Alignment and PRISM, CuMA outperforms dense PEFT baselines and semantic-only MoE adapters. It also transfers to held-out demographic groups with modest degradation. This makes it a practical template for models that must condition answers on the asker's cultural background.

Key points

  • On WVB with Qwen3-8B, CuMA reaches 50.64% accuracy, 31.50 Macro-F1 and 0.1876 EMD. LoRA reaches 40.06% accuracy and HydraLoRA 45.36% under full supervision.
  • In leave-one-cluster-out zero-shot tests, CuMA drops only about 2.5 points, to 48.10%. Baselines lose roughly 7.7 to 14.5 points, and LoRA falls to 32.40%.
  • Ablations show both routing signals matter:
    • Dropping demographic routing costs 3.56 accuracy points.
    • Dropping semantic routing costs 6.38 points.
    • Also removing the load-balancing loss raises EMD to 0.2657.
  • Top-k routing worked better than soft routing, because strict capacity separation is needed. The setup uses N = 8 experts.
  • The authors list several limitations:
    • The method relies on explicit demographic profiles, which may be missing or restricted for privacy reasons.
    • It inherits the coverage and biases of WVB, CA and PRISM.
    • It adds modest inference overhead.
  • Code is available at github.com/Throll/CuMA.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Southeast UniversityNanjing · China28 works · 354 citations🎓Southeast UniversityDhaka · Bangladesh8 works · 160 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Ao Sun, Xiaoyu Wang, Zhe Tan, Yu Li, Jiachen Zhu, Yuheng Jia and Shu Su. 2026. CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1265

@inproceedings{sun2026cuma,
  title = {{CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters}},
  author = {Ao Sun, Xiaoyu Wang, Zhe Tan, Yu Li, Jiachen Zhu, Yuheng Jia and Shu Su},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1265},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1265}
}

References

56 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. \#\# Conclusion We introduced C U MA, a framework that reformulates cultural alignment as a conditional capacity separation problem. By using demographic-aware routing, C U MA learns a Latent Cultural Topology to disentangle conflicting gradients and resolve Mean Collapse. Results across three benchma… find
  2. Muhammad Farid Adilazuarda, Sagnik Mukherjee, Pradhyumna Lavania, Siddhant Shivdutt Singh, Alham Fikri Aji, Jacki O’Neill, Ashutosh Modi, and Monojit Choudhury. 2024. Towards measuring and modeling “culture” in LLMs: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lan…
  3. Janice Ahn, Rishu Verma, Renze Lou, Di Liu, Rui Zhang, and Wenpeng Yin. 2024. Large language models for mathematical reasoning: Progresses and challenges. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages…
  4. Badr AlKhamissi, Muhammad ElNokrashy, Mai Alkhamissi, and Mona Diab. 2024. Investigating cultural alignment of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12404–12422, Bangkok, Thailand. Association…
  5. Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. 2024. Llemma: An open language model for mathematics. Preprint, arXiv:2310.10631.
  6. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, and 39 others. 202… — Evaluating Large Language Models Trained on Code
  7. Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 4302–4310, Red Hook, NY, USA. Curran Associa… find
  8. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.
  9. Esin Durmus, Karina Nguyen, Thomas I. Liao, Nicholas Schiefer, Amanda Askell, Anton Bakhtin, Carol Chen, Zac Hatfield-Dodds, Danny Hernandez, Nicholas Joseph, Liane Lovitt, Sam McCandlish, Orowa Sikder, Alex Tamkin, Janel Thamkul, Jared Kaplan, Jack Clark, and Deep Ganguli. 2024. Towards measuring t…
  10. Evan Frick. 2025. Reward modeling for human preferences. Master’s thesis, EECS Department, University of California, Berkeley, May. find
  11. Yao Fu and Mirella Lapata. 2022. Latent topology induction for understanding contextualized representations. Preprint, arXiv:2206.01512. — [arXiv:2206.01512](https://arxiv.org/abs/2206.01512) find
  12. Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2024. Large language models empowered agent-based modeling and simulation: a survey and perspectives. Humanities and Social Sciences Communications, 11(1):1259. find
  13. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  14. Yanggan Gu, Yuanyi Wang, Zhaoyi Yan, Yiming Zhang, Qi Zhou, Fei Wu, and Hongxia Yang. 2025. Infifpo: Implicit model fusion via preference optimization in large language models. Preprint, arXiv:2505.13878. — [arXiv:2505.13878](https://arxiv.org/abs/2505.13878) find
  15. Joseph Henrich, Steven J. Heine, and Ara Norenzayan. 2010. The weirdest people in the world? Behavioral and Brain Sciences, 33(2–3):61–83. find
  16. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. Preprint, arXiv:2106.09685.
  17. Ronald Inglehart and Christian Welzel. 2005. Modernization, Cultural Change, and Democracy: The Human Development Sequence. Cambridge University Press. find
  18. Liwei Jiang, Taylor Sorensen, Sydney Levine, and Yejin Choi. 2025. Can language models reason about individualistic human values and preferences? In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vienna, Austria. Association for Compu…
  19. Chengyi Ju, Weijie Shi, Chengzhong Liu, Jiaming Ji, Jipeng Zhang, Ruiyuan Zhang, Jiajie Xu, Yaodong Yang, Sirui Han, and Yike Guo. 2025. Benchmarking multi-national value alignment for large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 20042–20058. find
  20. Mehdi Khamassi, Marceau Nahon, and Raja Chatila. 2024. Strong and weak alignment of large language models with human values. Scientific Reports, 14(1):19399. find
  21. Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale. 2024. The prism alignment dataset: What participatory, representative and individualised human feedback reveals abo… find
  22. Ekaterina Kostina, Larisa Kretova, Raisa Teleshova, Anna Tsepkova, and Timur Vezirov. 2015. Universal human values: Cross-cultural comparative analysis. Procedia - Social and Behavioral Sciences, 214:1019– 1028. Worldwide trends in the development of education and academic research, Sofia, Bulgaria,… find
  23. Cheng Li, Mengzhuo Chen, Jindong Wang, Sunayana Sitaram, and Xing Xie. 2024a. Culturellm: Incorporating cultural differences into large language models. In Advances in Neural Information Processing Systems, volume 37, pages 84799–84838. Curran Associates, Inc.
  24. Dengchun Li, Yingzi Ma, Naizheng Wang, Zhengmao Ye, Zhiyuan Cheng, Yinghao Tang, Yan Zhang, Lei Duan, Jie Zuo, Cal Yang, and Mingjie Tang. 2024b. find
  25. Mixlora: Enhancing large language models fine-tuning with lora-based mixture of experts. Preprint, arXiv:2404.15159. Human Language Technologies (Volume 1: Long Papers), Mexico City, Mexico. Association for Computational Linguistics. find
  26. Jiyi Li. 2024. A comparative study on annotation quality of crowdsourcing and llm via label aggregation. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6525–6529. find
  27. Erik Miehling, Michael Desmond, Karthikeyan Natesan Ramamurthy, Elizabeth M. Daly, Pierre Dognin, Jesus Rios, Djallel Bouneffouf, and Miao Liu. 2025. Evaluating the prompt steerability of large language models. Preprint, arXiv:2411.12405. find
  28. Tianyi Li, Divya Sree, and Tatiana Ringenberg. 2025. Assessing crowdsourced annotations with LLMs: Linguistic certainty as a proxy for trustworthiness. In Proceedings of the 5th International Conference on Natural Language Processing for Digital Humanities, pages 191–201, Albuquerque, USA. Associati… find
  29. Juhyun Oh, Inha Cha, Michael Saxon, Hyunseung Lim, Shaily Bhatt, and Alice Oh. 2025. Culture is everywhere: A call for intentionally cultural evaluation. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 19156–19168, Suzhou, China. Association for Computational Linguist… find
  30. Ziyue Li and Tianyi Zhou. 2024. Your mixture-of-experts llm is secretly an embedding model for free. Preprint, arXiv:2410.10814. — [arXiv:2410.10814](https://arxiv.org/abs/2410.10814) find
  31. Haijiang Liu, Yong Cao, Xun Wu, Chen Qiu, Jinguang Gu, Maofu Liu, and Daniel Hershcovich. 2025. Towards realistic evaluation of cultural value alignment in large language models: Diversity enhancement for survey response simulation. Information Processing & Management, 62(4):104099. find
  32. Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. Dora: weight-decomposed low-rank adaptation. In Proceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org. find
  33. Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning v2: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 6… find
  34. Haoran Lu, Luyang Fang, Ruidong Zhang, Xinliang Li, Jiazhang Cai, Huimin Cheng, Lin Tang, Ziyu Liu, Zeliang Sun, Tao Wang, Yingchuan Zhang, Arif Hassan Zidan, Jinwen Xu, Jincheng Yu, Meizhi Yu, Hanqi Jiang, Xilin Gong, Weidi Luo, Bolun Sun, and 31 others. 2025. Alignment and safety in large language… — [arXiv:2507.19672](https://arxiv.org/abs/2507.19672) find
  35. Marlene Lutz, Indira Sen, Georg Ahnert, Elisa Rogers, and Markus Strohmaier. 2025. The prompt makes the person(a): A systematic evaluation of sociodemographic persona prompting for large language models. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 23212–23237, Suz… find
  36. Anemily Machina and Robert Mercer. 2024. Anisotropy is not inherent to transformers. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Fırat Öncel, Matthias Bethge, Beyza Ermis, Mirco Ravanelli, Cem Subakan, and Çağatay Yıldız. 202… find
  37. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744.
  38. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, volume 36, pages 53728–53741. Curran Associates, Inc.
  39. Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose opinions do language models reflect? In International Conference on Machine Learning, pages 29971–30004. PMLR.
  40. Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538. find
  41. Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. 2024. A roadmap to pluralistic alignment. Preprint, arXiv:2402.05070.
  42. Nicholas Sukiennik, Chen Gao, Fengli Xu, and Yong Li. 2025. An evaluation of cultural value alignment in llm. Preprint, arXiv:2504.08863. — [arXiv:2504.08863](https://arxiv.org/abs/2504.08863) find
  43. Yan Tao, Olga Viberg, Ryan S Baker, and René F Kizilcec. 2024. Cultural bias and cultural alignment of large language models. PNAS Nexus, 3(9):pgae346. find
  44. Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. 2024. Hydralora: An asymmetric lora architecture for efficient fine-tuning. In Advances in Neural Information Processing Systems, volume 37, pages 9565–9584. Curran Associates, Inc. find
  45. Siqi Wang, Zhengyu Chen, Bei Li, Keqing He, Min Zhang, and Jingang Wang. 2024a. Scaling laws across model architectures: A comparative analysis of dense and MoE models in large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5583–5595… find
  46. Yuhang Wang, Yanxu Zhu, Chao Kong, Shuyu Wei, Xiaoyuan Yi, Xing Xie, and Jitao Sang. 2024b. CDEval: A benchmark for measuring the cultural dimensions of large language models. In Proceedings of the 2nd Workshop on Cross-Cultural Considerations in NLP, pages 1–16, Bangkok, Thailand. Association for C… find
  47. Shaoyang Xu, Yongqi Leng, Linhao Yu, and Deyi Xiong. 2025. Self-pluralising culture alignment for large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers),… find
  48. Wanqi Xue, Bo An, Shuicheng Yan, and Zhongwen Xu. 2024. Reinforcement learning from diverse human preferences. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 5298–5306. International Joint Conferences on Artificial Intelligence Organizat… find
  49. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  50. 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. 2025a. Cultivating pluralism in algorithmic monoculture: The community alignm… — [arXiv:2507.09650](https://arxiv.org/abs/2507.09650) find
  51. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025b. Qwen3 embedding: Advancing text embedding and reranking through foundation models. Preprint, arXiv:2506.05176.
  52. Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2025c. The lessons of developing process reward models in mathematical reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 10495–10516, Vienn…
  53. Wenlong Zhao, Debanjan Mondal, Niket Tandon, Danica Dillion, Kurt Gray, and Yuling Gu. 2024. World-ValuesBench: A large-scale benchmark dataset for multi-cultural value awareness of language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Reso… find
  54. Xuandong Zhao, Will Cai, Tianneng Shi, David Huang, Licong Lin, Song Mei, and Dawn Song. 2025. Improving llm safety alignment with dual-objective optimization. In Proceedings of the 42nd International Conference on Machine Learning, ICML’25. JMLR.org. find
  55. Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Y. Zhao, Andrew Dai, Zhifeng Chen, Quoc Le, and James Laudon. 2022. Mixture-of-experts with expert choice routing. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY, US… find
  56. Xiaoxuan Zhu, Zhouhong Gu, Baiqian Wu, Suhang Zheng, Tao Wang, Tianyu Li, Hongwei Feng, and Yanghua Xiao. 2025. Toremi: Topic-aware data reweighting for dynamic pre-training data selection. Preprint, arXiv:2504.00695. — [arXiv:2504.00695](https://arxiv.org/abs/2504.00695) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sun-2026-cuma-aligning-llms-sparse-cultural/