Bias control in large language models

Develop robust methods to control and mitigate negative biases in large language models in order to improve safety and reliability across applications that rely on these models.

Background

The paper notes that LLMs can inherit negative racial, gender, and other biases from their training data. Because training corpora are vast, heterogeneous, and often insufficiently documented, LLMs may encode harmful or undesirable societal biases that pose safety risks.

While stricter filtering of training content and better transparency into training datasets may help, the authors explicitly state that controlling such biases remains an open problem for LLMs, underscoring a need for principled, effective mitigation techniques.

References

Stricter filtering of training content and better transparency into a training sets data will provide better safety but bias-control is likely to be a longstanding open problem of LLMs.

— Evolving Code with A Large Language Model  (2401.07102 - Hemberg et al., 2024) in Section 2, Background: Large Language Models (Fourth paragraph on bias)

Developing robust and scalable mitigation strategies remains an open challenge.

— Hidden Threat in Synthetic Data: Covert Targeted Bias Injection through Benign Text  (2608.30619 - Cho et al., 31 Aug 2026) in Limitations

For arbitrary natural-language secrets, we do not know the set of all strings to which generation needs to be invariant, so it is unclear how to train a model to generate text in this way.

— Inadvertent Context Leakage in Language Models  (2608.19857 - Fairoze et al., 20 Aug 2026) in Section 6, paragraph “Toward mitigation.”

Recent work continues to find such effects in modern MT architectures, suggesting that default gender preferences remain an open problem even as model capabilities improve \citep{manna-etal-2026-gender}.

— Benchmarking Gender Bias in Machine Translation Evaluation Metrics across Occupations  (2609.21490 - Mastromichalakis et al., 18 Sep 2026) in Section 2, Related Work

Overall, research on English hate and OLD showed a strong move toward transformer-based models, with F1-scores approaching 90% on well-curated datasets, while issues of bias, interpretability, and robustness remain open challenges.

— An Explainable DistilBERT-BiLSTM-Attention Framework for Binary and Multi-Class Hate Speech Detection  (2609.28703 - Zia et al., 23 Sep 2026) in Section 2.1, Hate Speech Detection