Vishnu Kabir Chhabra

Restoring General and Improving Domain-Specific Safety in Fine-Tuned LLMs Without Retraining

According to the abstract, Chhabra and colleagues propose SPECTRA, a training-free method that copies a base model's refusal direction into a fine-tuned model's weights through a low-rank update, aiming to restore refusals lost after benign fine-tuning.