Steering LLMs Towards Unbiased Responses: A Causality-Guided Debiasing Framework

• DOI: 10.48550/arxiv.2403.08743
• Publication Date: 2024-03-13
• Authors: Jingling Li; Zeyu Tang; Xiaoyu Liu; Peter Spirtes; Kun Zhang; Liu Leqi; Yang Liu
• Keywords: Debiasing; Causality

Large language models (LLMs) can easily generate biased and discriminative responses. As LLMs tap into consequential decision-making (e.g., hiring healthcare), it is of crucial importance to develop strategies mitigate these biases. This paper focuses on social bias, tackling the association between demographic information LLM outputs. We propose a causality-guided debiasing framework that utilizes causal understandings (1) data-generating process training corpus fed LLMs, (2) internal reasoning inference, guide design prompts for outputs through selection mechanisms. Our unifies existing de-biasing prompting approaches such as inhibitive instructions in-context contrastive examples, sheds light new ways by encouraging bias-free reasoning. strong empirical performance real-world datasets demonstrates our provides principled guidelines even with only black-box access.