This research introduces a novel method for making large language models securely forget specific information by steering their internal representations.

• Representation steering in intermediate layers reduces token confidence, causing models to generate incorrect responses
• The approach offers theoretical explanations for why representation manipulation leads to effective unlearning
• Method demonstrates robustness against adversarial jailbreak attacks
• Provides critical insights for implementing selective knowledge removal in AI systems

Why it matters: Controlled forgetting is essential for privacy compliance, removing harmful content, and protecting LLMs against security exploits while maintaining performance on desired tasks.

On Effects of Steering Latent Representation for Large Language Model Unlearning