UNLEARN Efficient Removal of Knowledge in Large Language Models

Name: UNLEARN Efficient Removal of Knowledge in Large Language Models
Uploaded: 2026-10-17T16:11:06.849162+00:00
Duration: 298 s
Description: UNLEARN proposes a method to dynamically forget specific knowledge in large language models by using subspace methods. It identifies and removes task-specific knowledge by orthogonalizing subspaces, ensuring that other general knowledge remains intact. This allows for precise unlearning, achieving 96% forgetting while preserving model performance on other tasks.

UNLEARN proposes a method to dynamically forget specific knowledge in large language models by using subspace methods. It identifies and removes task-specific knowledge by orthogonalizing subspaces, ensuring that other general knowledge remains intact. This allows for precise unlearning, achieving 96% forgetting while preserving model performance on other tasks.

Open the full episode