📝 SummarXiv

Abstract

Moral alignment has emerged as a widely adopted approach for regulating the behavior of pretrained language models (PLMs), typically through fine-tuning or model editing on curated datasets. However, this process often comes at the cost of degraded downstream task performance. Prior studies commonly aim to achieve a performance trade-off by encouraging PLMs to selectively forget stereotypical knowledge through carefully designed fairness objectives, while preserving their helpfulness. In this short paper, we investigate the underlying mechanisms of the performance trade-off in the context of mitigating gender stereotypes, through the lens of forgetting and the fairness objective. Our analysis reveals the limitations of current fairness objective in achieving trade-off by demonstrating that: (1) downstream task performance is primarily driven by the overall forgetting level; (2) selective forgetting of stereotypes tends to increase overall forgetting; and (3) general solutions for mitigating forgetting are ineffective at reducing overall forgetting and fail to improve downstream task performance.