Measuring and mitigating stereotypes in large language models

Completely ‘debiasing’ a language model is, in many respects, an impossible task. Language is inherently contextual, and individuals interpret text differently based on their lived experiences, social positions, and cultural backgrounds. This subjectivity makes stereotypes difficult to define, measure, and mitigate in a consistent way. This thesis engages directly with these challenges, measuring which stereotypes are encoded in Large Language Models (LLMs) and exploring practical approaches for reducing representational harm despite the absence of a single ground truth. This thesis introduces GlobalBias, a dataset that incorporates a wide range of intersectional demographic groups, alongside a methodology and novel evaluation metric that surface stereotypical associations without pre-defining which stereotypes should be linked to particular identities. By avoiding fixed stereotype labels, this approach reveals which associations emerge from model behaviour in both open-source models (via perplexity) and closed-source models (via generative outputs). The analysis centres on the harms these patterns cause, particularly for historically oppressed and marginalised groups. The thesis then shifts from analysis to intervention. We present the first systematic application of steering vectors to social bias reduction, showing that targeted interventions at the level of model activations can significantly reduce stereotypical behaviour. Across multiple benchmarks, steering vectors consistently outperform prompting, Self-Debias, and finetuning, achieving substantial bias reduction while preserving overall model performance. In addition, we introduce Dialz, an open source toolkit that enables reproducible training and evaluation of steering vectors, alongside automated methods for selecting optimal intervention layers and coefficients. Taken together, this thesis contributes a novel dataset, methodological frameworks, empirical results, and practical tools for understanding and mitigating bias in LLMs. By combining rigorous measurement and a critical engagement with the wider societal implications of stereotypical language, this work advances approaches to developing language models that are more transparent and aligned with diverse human experiences.