level: research
large language models are used in security-critical areas like healthcare, finance, and education, but they cannot forget. sensitive personal data, copyrighted content, and hazardous knowledge stay encoded in their parameters long after deployment. this creates risks such as data extraction, jailbreak attacks, membership inference, and breaking privacy rules. real incidents include chatbots repeating private information and making up legal citations that cause real harm. retraining these huge models on cleaned data is too expensive, and knowledge is spread across parameters, not stored in one place.
the survey looks at unlearning methods that aim to remove specific knowledge without full retraining. these include gradient-based updates, parameter editing, and in-context unlearning. each approach has trade-offs in effectiveness, efficiency, and side effects like degrading other model abilities. the paper also covers evaluation metrics, such as testing if unlearned knowledge can still be extracted or if model utility drops. challenges include verifying that unlearning is complete and dealing with knowledge that is entangled across the model.
emerging threats make unlearning urgent. attackers can probe models to recover supposedly deleted information, and new regulations require data deletion. the survey connects unlearning to cyber defense, showing how forgetting dangerous capabilities—like generating exploit code or revealing system vulnerabilities—can reduce attack surfaces. it also discusses open problems, such as scalable unlearning for very large models and defending against adaptive adversaries who try to reverse the unlearning process.
why it matters: unlearning helps remove private or dangerous knowledge from ai models, reducing security and privacy risks without costly retraining.