Researchers have developed FairNVT, a lightweight debiasing framework for transformer-based encoders. It aims to improve prediction fairness by reducing sensitive-attribute information in the representation used by downstream classifiers. This is achieved through adapters that learn task-relevant and sensitive embeddings, calibrated Gaussian noise applied to the sensitive embedding, and orthogonality constraints and fairness regularization. The approach has been tested on three datasets spanning vision and language tasks, resulting in improved fairness metrics such as demographic parity difference and equalized odds, while maintaining competitive task performance. --- Why it matters: This matters because current AI models often perpetuate biases present in the data they're trained on, leading to unfair outcomes. FairNVT's approach provides a way to mitigate these biases without sacrificing task performance, which is crucial for applications like image classification and natural language processing where fairness is essential. Source: https://arxiv.org/abs/2604.16780