GradShield: Alignment Preserving Finetuning
Researchers have developed GradShield, a new method to prevent large language models from becoming misaligned after fine-tuning. The technique identifies and removes harmful data points before they can corrupt the model's alignment by calculating a Finetuning Implicit Harmfulness Score (FIHS) for each data point. Experiments show GradShield effectively maintains model utility while keeping the attack success rate below 6%, outperforming existing baseline methods. AI
IMPACT This method could enable safer deployment of fine-tuned LLMs by preventing the introduction of harmful behaviors.