Minseong Kim

Rethinking LLM Fine-Tuning via Weight Space Reparameterization: Preserving Safety during Downstream Adaptation

According to the abstract, Minseong Kim, Jongbok Won, Jeesup Park and colleagues propose WSR-Tune, a fine-tuning method that reparameterizes weights so safety information concentrates in a few directions, then freezes them while other directions adapt to the task. They aim to keep safety when models are fine-tuned.