Elamur Research⌕
☰
home / papers / 2026 / An Investigation of Linguistic Biases in LLM-Based Recommendations

An Investigation of Linguistic Biases in LLM-Based Recommendations

Paper · 2026 · Nitin Venkateswaran, Jason Ang, Deep Adhikari et al.

Summary

The paper asks whether LLM recommenders change their suggestions when the same request is phrased in a different dialect. The authors wrote 30 questions each for restaurant and product scenarios. They rendered them in Southern American English and Indian English with the rule-based Multi-VALUE toolkit, and in code-switched Hindi-English through ChatGPT translation followed by manual checking. Each prompt contained a freshly sampled, balanced list of candidate names: American vs. Indian restaurants from the Yelp Open Dataset, or seven Walmart product categories. Six instruction models from the llama-3.1, gpt-oss and mistral families were asked, with no examples or history (a cold-start setting), for a top-20 pick. The authors counted the picks per cuisine or category and analysed them with mixed-effects regression, likelihood ratio tests and pairwise comparisons of estimated marginal means. Dialect alone shifted the recommendations by a statistically significant amount. The llama-3.1 models reacted most strongly, gpt-oss least, and model size showed no consistent effect. The result exposes a fairness and user-autonomy risk that has received little study.

Key points

  • Candidate pools came from two datasets:
    • The Yelp Open Dataset, filtered to 18,203 American and 1,086 Indian restaurants.
    • The Walmart product dataset, spanning seven categories, listed alphabetically: Beauty, Clothing, Exercise, Health, Home, Personal Care, and Sports & Outdoors.
  • Prompt lists and filtering:
    • Each prompt list held 100 names per type or category, drawn by uniform random sampling.
    • A new list was drawn for every seed, question and dialect, with 20 seeds per question.
    • Only replies containing exactly 20 names were kept.
  • Six models were tested: llama-3.1 8B/70B, gpt-oss 20B/120B, and mistral-7B-instruct vs. mistral-small-3.1 (24B).
  • llama-3.1-70B was the most sensitive to code-switched prompts. With them it recommended more Indian restaurants and shifted counts in four of the seven product categories.
  • Beauty picks increased under Indian English prompts for the larger models. Home picks increased under code-switched prompts for the smaller models. Model size by itself showed no general bias trend.

Interesting ideas

  • I1. The study builds meaning-preserving variants of each prompt by applying rule-based dialect transformations, then treats any change in output as a defect, which makes it a cheap invariance probe for any system.
  • I2. The same dialect rewrites amount to a controlled syntactic perturbation with known intent. That makes them useful for stress-testing analysis components on non-standard English as well as for checking final outputs.
  • I3. The model sees a different balanced candidate list on every seed, so conclusions cannot depend on one lucky or unlucky fixed sample.
  • I4. Prompts list bare names only, with every other attribute removed, so that any bias can be traced to a single source.
  • I5. Response counts are analysed with mixed-effects models, likelihood ratio tests and pairwise estimated marginal means rather than judged from boxplots. This separates real group effects from per-question and per-seed noise.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (answer): Run school tests and fable QA through Multi-VALUE dialect rewrites and flag any question whose answer flips, since the meaning has not changed.
  • I2 (tree): Parse each original sentence and its dialect variant, then compare the trees on the shared content words to find constructions where the perceptron parser or gparse breaks on dialect morphosyntax.
  • I3 (learn): Draw fresh, balanced evaluation subsets per seed when comparing parser or teacher configurations, instead of reusing one frozen dev slice.
  • I4 (learn): When auditing LLM teacher tiers, strip the context down step by step to tell whether a teacher's labelling shift comes from the sentence form or from side information.
  • I5 (learn): Use the same statistical setup, with questions and seeds as random effects, to decide whether a new repair rule or training change truly improves accuracy.

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

🎓University of FloridaGainesville · United States33 works · 364 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

Cite

Cite as Nitin Venkateswaran, Jason Ang, Deep Adhikari and Tarun Krishna Dasari. 2026. An Investigation of Linguistic Biases in LLM-Based Recommendations. arXiv:2604.25456.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2604.25456 · arXiv 2604.25456

@misc{venkateswaran2026investigation,
  title = {{An Investigation of Linguistic Biases in LLM-Based Recommendations}},
  author = {Nitin Venkateswaran, Jason Ang, Deep Adhikari and Tarun Krishna Dasari},
  year = {2026},
  doi = {10.48550/arxiv.2604.25456},
  eprint = {2604.25456},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2604.25456}
}

References

25 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, and 1 others. 2025. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925. ↗
  2. Alexandre Andre, Gauthier Roy, Eva Dyer, and Kai Wang. 2025. Revealing potential biases in llm-based recommender systems in the cold start setting. arXiv preprint arXiv:2508.20401. ↗
  3. Sofia Bonicalzi, Mario De Caro, and Benedetta Giovanola. 2023. Artificial intelligence and autonomy: on the ethical dimension of recommender systems. Topoi, 42(3):819–832. ↗
  4. Nicholas Deas, Jessica Grieser, Shana Kleiner, Desmond Patton, Elsbeth Turcan, and Kathleen McKeown. 2023. Evaluation of African American language bias in natural language generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6805– 6824, Singapore. Association for Computational Linguistics. find
  5. and William El Sayed. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825. ↗
  6. Yashar Deldjoo. 2025. Understanding biases in chatgpt-based recommender systems: Provider fairness, temporal stability, and recency. ACM Trans. Recomm. Syst., 4(2). ↗
  7. Russell V. Lenth and Julia Piaskowski. 2026. emmeans: Estimated Marginal Means, aka Least-Squares Means. R package version 2.0.3. find
  8. Jan Malte Lichtenberg, Alexander Buchholz, and Pola Schwöbel. 2024. Large language models as recommender systems: A study of popularity bias. ↗
  9. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186. ↗
  10. Fangru Lin, Shaoguang Mao, Emanuele La Malfa, Valentin Hofmann, Adrian de Wynter, Jing Yao, Si-Qing Chen, Michael J Wooldridge, and Furu Wei. 2024. One language, many gaps: Evaluating dialect fairness and robustness of large language models in reasoning tasks. find
  11. Hanpei Fang, Sijie Tao, Nuo Chen, Kai-Xin Chang, and Tetsuya Sakai. 2025. Do large language models favor recent content? a study on recency bias in llm-based reranking. In Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, SIGIR-AP 2025, page 85–94, New York, NY, USA. Association for Computing Machinery. find
  12. PromptCloud. 2020. Walmart product reviews dataset. https://www.kaggle.com/datasets/promptcloud/ walmart-product-reviews-dataset. Accessed: 2026-04-15. ↗
  13. R Core Team. 2024. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. find
  14. Eve Fleisig, Genevieve Smith, Madeline Bossi, Ishita Rustagi, Xavier Yin, and Dan Klein. 2024. Linguistic bias in ChatGPT: Language models reinforce dialect discrimination. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 13541–13564, Miami, Florida, USA. Association for Computational Linguistics.
  15. Tianshu Shen, Jiaru Li, Mohamed Reda Bouadjenek, Zheda Mai, and Scott Sanner. 2023. Towards understanding and mitigating unintended biases in language model-driven conversational recommendation. Information Processing & Management, 60(1):103139. ↗
  16. John Fox and Sanford Weisberg. 2019. An R Companion to Applied Regression, third edition. Sage, Thousand Oaks CA. find
  17. Mengting Wan, Jianmo Ni, Rishabh Misra, and Julian McAuley. 2020. Addressing marketing bias in product recommendations. In Proceedings of the 13th international conference on web search and data mining, pages 618–626. find
  18. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. ↗
  19. Bohao Wang, Jiawei Chen, Feng Liu, Changwang Zhang, Jun Wang, Canghong Jin, Chun Chen, and Can Wang. 2025. Does llm focus on the right words? diagnosing language bias in llm-based recommenders. arXiv preprint arXiv:2510.10978. ↗
  20. Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. In Advances in Information Retrieval: 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24–28, 2024, Proceedings, Part II, page 364–381, Berlin, Heidelberg. Springer-Verlag. find
  21. Ke Xu, Shera Potka, and Alex Thomo. 2025. Gender and race bias in consumer product recommendations by large language models. In International Conference on Advanced Information Networking and Applications, pages 245–258. Springer. find
  22. Yelp Inc. 2023. Yelp open dataset. https://www.yelp.com/dataset. Accessed: 2026-04-15. ↗
  23. Wenyue Hua, Yingqiang Ge, Shuyuan Xu, Jianchao Ji, Zelong Li, and Yongfeng Zhang. 2024. Up5: Unbiased foundation model for fairness-aware recommendation. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1899–1912. ↗
  24. Yuhui Zhang, Hao Ding, Zeren Shui, Yifei Ma, James Zou, Anoop Deoras, and Hao Wang. 2021. Language models as recommender systems: Evaluations and limitations. find
  25. Caleb Ziems, William Held, Jingfeng Yang, Jwala Dhamala, Rahul Gupta, and Diyi Yang. 2023. Multi-VALUE: A framework for cross-dialectal English NLP. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 744–768, Toronto, Canada. Association for Computational Linguistics. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix,

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/venkateswaran-2026-investigation-linguistic-biases-llm-based/