Causal prompting method reduces biases in large language models, suggesting effective mitigation strategies.
Key Points
Causal prompting achieves effective bias mitigation in large language models by utilizing structural causal models to uncover underlying relationships.
Experimental results demonstrate that this method improves performance across three natural language processing datasets, indicating strong applicability and robustness.
Assessment uses contrastive learning to fine-tune the encoder's representation, enhancing the accuracy of causal effect estimations for given prompts and outputs, facilitating better alignment with LLMs' responses.