Exclusive Unlearning: Achieving Safe Alignment of Large Language Models via 'Retention-Based Forgetting'

The study proposes the Exclusive Unlearning (EU, Retention-Based Forgetting) method. By reversing the traditional machine forgetting approach—specifying content to retain and forgetting all other information—it achieves comprehensive elimination of diverse harmful content while preserving professional capabilities in specific domains (e.g., medicine, mathematics), providing a new path for the safe alignment of Large Language Models (LLMs).

LLMs are widely used in fields such as healthcare and education, but the risk of generating harmful content is prominent. Traditional safety alignment methods (SFT, RLHF) struggle to cover all harmful scenarios, are easily bypassed by jailbreak attacks, and excessive safety measures may inadvertently damage useful capabilities. Traditional machine forgetting requires enumerating forgetting targets one by one, leading to high computational costs when dealing with diverse harmful content and inability to prevent emergent harmfulness.

The core of EU is a paradigm reversal of 'retain whitelist + forget the rest': no need to list harmful items, only specify the content to retain. The technical steps include: 1) Constructing a retention dataset without harmful content; 2) Adopting an objective function of 'minimizing the likelihood of non-retained data + moderately fitting retained data'; 3) Tuning hyperparameters to balance safety and usefulness.

Models processed by EU are extremely robust against jailbreak attacks (attack success rate is nearly zero) because the model has forgotten the knowledge to generate harmful content. In the medical field, it retains diagnostic/treatment knowledge while being 'ignorant' of harmful medical queries, and its performance in medical exams is comparable to the original model. In the mathematical field, it retains problem-solving abilities and forgets abusive content (such as cracking encryption).

Compared to RLHF/SFT: EU fundamentally removes harmful capabilities instead of only training to refuse answers. Compared to traditional machine forgetting: EU does not require forgetting item by item, has strong scalability, can prevent emergent harmfulness, and achieves extensive forgetting with one training session.

Limitations: Difficulty in designing retention datasets (need to be comprehensive and without loopholes), possible reduction in general capabilities, and high computational costs. Future directions: Intelligent construction of retention datasets, efficient training algorithms, combination of EU with other safety technologies, expansion to multimodal models, and improvement of evaluation systems.

EU is an important paradigm shift in the field of machine forgetting, providing a new path for building safe and trustworthy LLMs. Against the backdrop of increasingly complex harmful content and jailbreak attacks, it has theoretical and practical value. Although there is room for improvement, it opens up a new direction for LLM safe alignment research.