Psychometric utility-preserving embedding debiasing (PUPD): A dynamic framework for bias-aware and culturally inclusive language models

Abstract Psychometric Utility-Preserving Embedding Debiasing (PUPD) is proposed as a dynamic framework for bias-aware and culturally inclusive language models that explicitly balances bias mitigation with semantic utility preservation. Conventional debiasing methods often disrupt task-critical embeddings or fail to account for intersectional biases, limiting their applicability in educational and culturally sensitive contexts. The proposed framework reformulates the embedding layer through two novel modules: the Psychometric Bias-Utility Quantifier (PBUQ), which evaluates bias-utility trade-offs across embedding subspaces, and the Heterogeneous Preference Optimization Network (HPON), which dynamically adjusts embedding updates to minimize bias while preserving utility. PBUQ quantifies intersectional biases through a learned attention mechanism over demographic attributes, whereas HPON frames the debiasing process as a multi-objective optimization problem with adaptive weighting. The framework integrates seamlessly with transformer architectures, refining token embeddings prior to their entry into self-attention blocks. Additionally, PUPD employs a gated recurrent unit (GRU) to model temporal bias-utility dynamics and a lightweight transformer encoder for cross-metric bias analysis. Experiments on three text-based fairness benchmarks — BiasBios, StereoSet, and IntersectionalBias — demonstrate that PUPD outperforms post-hoc debiasing techniques by maintaining higher semantic coherence throughout bias reduction.