Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought

arXiv:2403.05518v1 Announce Type: new
Abstract: While chain-of-thought prompting (CoT) has the potential to improve the explainability of language model reasoning, it can systematically misrepresent the factors influencing models' behavior--for example, rationalizing answers in line with a user's opinion without mentioning this bias. To mitigate this biased reasoning problem, we introduce bias-augmented consistency training (BCT), an unsupervised fine-tuning scheme that trains models to give consistent reasoning across prompts with and without biasing features. We construct a suite testing nine forms of …

abstract arxiv behavior bias cs.ai cs.cl example explainability language language model line opinion prompting reasoning thought training type