1 article
New reasoning-based LLM unlearning method cuts model bias 40% by surgically removing unsafe knowledge—without full retraining.