Elamur Research⌕
☰
home / papers / 2026 / Large Language Models Are Still Misled by Simple Bias Ensembles

Large Language Models Are Still Misled by Simple Bias Ensembles

Paper · 2026 · Zhouhao Sun, Zhiyuan Kan, Xiao Ding et al.

Summary

Recent LLMs handle individual shortcut cues better than before. This paper argues that the real weakness appears when several such cues occur together in one input, which is common in real-world text. Most existing bias benchmarks inject only one bias per sample, and a few inject two. The authors therefore built MB-Ben, a natural language inference (NLI) benchmark in which every example carries five simple biases that all push the model toward the same label.

They first confirmed which known NLI biases still affect GPT-5.1, using MNLI samples, and determined the label each bias favours. They then wrote premise–hypothesis pairs, including counterfactual neutral and contradiction cases that are rare in existing corpora, so that the stacked cues point toward "entailment" while the gold label may differ.

Mainstream LLMs and existing debiasing methods both performed markedly worse on MB-Ben than on single-bias data. Larger models and slow-thinking modes did not offer a resource-efficient fix. The work shows that stacked biases remain an open robustness problem, with implications for high-stakes uses such as clinical and legal analysis.

Key points

  • The authors catalogue eight NLI bias features and the label each one favours. Five that all favour "entailment" are combined in every MB-Ben item:
    • the hypothesis is more than five words shorter than the premise;
    • lexical overlap is above 0.8;
    • BERTScore is above 0.88;
    • a speculative word (e.g., "might") is present;
    • a male subject has a male-stereotyped occupation.
  • The semantic-similarity thresholds were chosen so that roughly 15% of MNLI falls into each extreme. High similarity means BERTScore above 0.88, and low similarity means below 0.83.
  • Bias probing used 3,000 label-balanced MNLI samples per feature with GPT-5.1-2025-11-13. On high-similarity data, GPT-5.1 predicted "entailment" 40.7% of the time, against an expected 33.3%. On low-similarity data, it predicted "neutral" 44.0% of the time.
  • The authors also built a separate setting with conflicting polarities, combining high semantic similarity with low lexical overlap.
  • Existing debiasing methods failed in multi-bias scenarios, including domain-context calibration, batch calibration, causal-guided active learning and UniBias. The dataset is released at github.com/spirit-moon-fly/MB-Ben.

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

🎓Harbin Institute of TechnologyHarbin · China158 works · 28,841 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 Zhouhao Sun, Zhiyuan Kan, Xiao Ding, Li Du, Bibo Cai, Yang Zhao, Bing Qin and Ting Liu. 2026. Large Language Models Are Still Misled by Simple Bias Ensembles. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.971

@inproceedings{sun2026large,
  title = {{Large Language Models Are Still Misled by Simple Bias Ensembles}},
  author = {Zhouhao Sun, Zhiyuan Kan, Xiao Ding, Li Du, Bibo Cai, Yang Zhao, Bing Qin and Ting Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.971},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.971}
}

References

28 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  2. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  3. Yu Fei, Yifan Hou, Zeming Chen, and Antoine Bosselut. 2023. Mitigating label biases for in-context learning. In Proceedings Of The 61St Annual Meeting Of The Association For Computational Linguistics (Acl 2023): Long Papers, Vol 1, pages 14014–14031. Assoc Computational Linguistics-Acl.
  4. Panatchakorn Anantaprayoon, Masahiro Kaneko, and Naoaki Okazaki. 2024. Evaluating gender bias of pre-trained language models in natural language inference by considering all labels. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluat… find
  5. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technol…
  6. Anastasiia Klimashevskaia, Dietmar Jannach, Mehdi Elahi, and Christoph Trattner. 2024. A survey on popularity bias in recommender systems. User Modeling and User-Adapted Interaction, 34(5):1777–1834. find
  7. Tianhe Lin, Jian Xie, Siyu Yuan, and Deqing Yang. 2025. Implicit reasoning in transformers is reasoning through shortcuts. arXiv preprint arXiv:2503.07604. find
  8. Bo Lv, Nayu Liu, Yang Shen, Xin Liu, Ping Luo, and Yue Yu. 2025. Whether llms know if they know: Identifying knowledge boundaries via debiased historical in-context learning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 19516–19528. find
  9. Marta Marchiori Manerba, Karolina Stanczak, Riccardo Guidotti, and Isabelle Augenstein. 2024. Social bias probing: Fairness benchmarking for language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14653–14671. find
  10. Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428–3448.
  11. Nick Mckenna, Tianyi Li, Liang Cheng, Mohammad Hosseini, Mark Johnson, and Mark Steedman. 2023. Sources of hallucination by large language models on inference tasks. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2758–2774.
  12. Roberto Navigli, Simone Conia, and Björn Ross. 2023. Biases in large language models: Origins, inventory and discussion. ACM Journal of Data and Information Quality. find
  13. OpenAI. 2025. Gpt-5 system card. find
  14. Sara Rajaee, Yadollah Yaghoobzadeh, and Mohammad Taher Pilehvar. 2022. Looking at the overlooked: An analysis on the word-overlap bias in natural language inference. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10605–10616.
  15. Timo Schick, Sahana Udupa, and Hinrich Schütze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in nlp. Transactions of the Association for Computational Linguistics, 9:1408– 1424.
  16. Julius Steen and Katja Markert. 2024. Bias in news summarization: Measures, pitfalls and corpora. In Findings of the Association for Computational Linguistics ACL 2024, pages 5962–5983. find
  17. Zechen Sun, Yisheng Xiao, Juntao Li, Yixin Ji, Wenliang Chen, and Min Zhang. 2024a. Exploring and mitigating shortcut learning for generative large language models. In Proceedings of the 2024 joint international conference on computational linguistics, language resources and evaluation (LREC-COLING… find
  18. Zhouhao Sun, Li Du, Xiao Ding, Yixuan Ma, Yang Zhao, Kaitao Qiu, Ting Liu, and Bing Qin. 2024b. Causal-guided active learning for debiasing large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14455– 14469. find
  19. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  20. Qwen Team. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  21. Adina Williams, Nikita Nangia, and Samuel R Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In NAACL-HLT.
  22. Shuo Yang, Bardh Prenkaj, and Gjergji Kasneci. 2025. Razor: Sharpening knowledge by cutting bias with unsupervised text rewriting. In Proceedings of the AAAI Conference on Artificial Intelligence. find
  23. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In Proceedings of the 8th International Conference on Learning Representations. find
  24. Han Zhou, Xingchen Wan, Lev Proleev, Diana Mincu, Jilin Chen, Katherine Heller, and Subhrajit Roy. 2024a. Batch calibration: Rethinking calibration for in-context learning and prompt engineering. In Proceedings of the Twelfth International Conference on Learning Representations. find
  25. Hanzhang Zhou, Zijian Feng, Zixiao Zhu, Junlang Qian, and Kezhi Mao. 2024b. Unibias: Unveiling and mitigating llm bias through internal attention and ffn manipulation. In Advances in Neural Information Processing Systems, volume 37, pages 102173–102196. Curran Associates, Inc. find
  26. Yuhang Zhou, Paiheng Xu, Xiaoyu Liu, Bang An, Wei Ai, and Furong Huang. 2024c. Explore spurious correlations at the concept level in language models for text classification. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 478–… find
  27. Yuqing Zhou, Ruixiang Tang, Ziyu Yao, and Ziwei Zhu. 2024d. Navigating the shortcut maze: A comprehensive analysis of shortcut learning in text classification by language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 2586–2614. find
  28. Premise Hypothesis N 1 is a P 1 , V 1 . N 1 is a P 1 . He V 1 . He S 1 V 2 . He S 1 V 2 . N 1 , a P 1 by trade, V 1 . N 1 , a P 1 by trade. he V 1 . He S 1 V 2 . He S 1 V 2 . N 1 works as a P 1 , V 1 . N 1 works as a P 1 . He V 1 . He S 1 V 2 . He S 1 V 2 . N 1 , recognized as a P 1 , V 1 . N 1 is r… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sun-2026-large-language-models-are-still/