The question-answering (QA) capabilities of foundation models are highly sensitive to prompt variations, rendering their performance susceptible to superficial, non-meaning-altering changes. This vulnerability often stems from the model's preference or bias towards specific input characteristics, such as option position or superficial image features in multi-modal settings. We propose to rectify this bias directly in the model's internal representation. Our approach, SteerFair, finds the bias direction in the model's representation space and steers activation values away from it during inference. Specifically, we exploit the observation that bias often adheres to simple association rules, such as the spurious association between the first option and correctness likelihood. Next, we construct demonstrations of these rules from unlabeled samples and use them to identify the bias directions. We empirically show that SteerFair significantly reduces instruction-tuned model performance variance across prompt modifications on three benchmark tasks. Remarkably, our approach surpasses a supervised baseline with 100 labels by an average of 10.86% accuracy points and 12.95 score points and matches the performance with 500 labels.

Large language models (LLMs) have garnered significant attention for their remarkable performance in a continuously expanding set of natural language processing tasks. However, these models have been shown to harbor inherent societal biases, or stereotypes, which can adversely affect their performance in their many downstream applications. In this paper, we introduce a novel, purely prompt-based approach to uncover hidden stereotypes within any arbitrary LLM. Our approach dynamically generates a knowledge representation of internal stereotypes, enabling the identification of biases encoded within the LLM's internal knowledge. By illuminating the biases present in LLMs and offering a systematic methodology for their analysis, our work contributes to advancing transparency and promoting fairness in natural language processing systems.

In the paper, we propose attention-based summarized post-hoc explanations for detection and identification of bias in data. We propose a global explanation and introduce a step-by-step framework on how to detect and test bias. Then, the bias is evaluated with a proposed counterfactual approach to bias insertion. Because removing the unwanted bias is often a complicated and tremendous task, we automatically insert it, instead. We validate our results on the example of the skin lesion dataset. Using the method, we successfully identified and confirmed part of the possible bias-causing artifacts in dermoscopy images. We confirmed that the commonplace black frames in the training dataset images have a strong influence on the Convolutional Neural Network's prediction. After artificially adding a black frame to all images, around 22% of them changed the prediction from benign to malignant. We have shown that bias detection is an important step of making more robust models, and we discuss how to improve them