This research evaluates whether unlearning techniques truly remove sensitive information from language model weights or merely hide access to it.

• Introduces an adversarial evaluation method to test information removal from model weights
• Focuses specifically on harmful capabilities like cyber-attacks and bioweapon creation
• Addresses a critical gap in understanding LLM safety mechanisms
• Provides insights for developing more effective unlearning methods

For security professionals, this research is vital as it helps distinguish between genuine safety improvements versus superficial barriers that determined attackers could bypass.

Do Unlearning Methods Remove Information from Language Model Weights?