Luxi He

Suppressing Misaligned Behaviors in Language Models, Durably

According to the abstract, a five-author team including Pengcheng Jiang and Jifan Zhang builds two model organisms with deliberate misalignment and compare methods for suppressing the behavior: fine-tuning, GRPO and unlearning. They judge each on immediate suppression, durability under further training and preserved capability, and introduce a new method.