The authors identified benign instructions that aligned models were initially inclined to refuse.
Benign fine-tuning can teach aligned models to answer harmful requests
Experiments show that harmless supervision can weaken safety when it pushes models from refusal toward compliance without offering an easy shortcut.
Independent
Jesson Wang · Zhanhao Hu · Minghao Lin · Yuchen Wu · Aylin Caliskan · Mengyuan Li
Research Digest··2 min read
Wang and colleagues show that fine-tuning data need not contain harmful content to erode an aligned language model’s refusal behavior.
§