Researchers have developed a new framework called ADU to help large language models 'forget' sensitive information. Unlike previous methods that simply erase or penalize target outputs, ADU decouples attention pathways to suppress unwanted knowledge while preserving linguistic structure. The authors claim their approach achieves the strongest performance on two benchmarks and reduces side effects in benign contexts by 92.9% compared to existing methods. --- Why it matters: This matters because large language models often struggle to balance forgetting sensitive information with retaining useful knowledge, which is crucial for addressing privacy regulations and safety concerns. Source: https://arxiv.org/abs/2608.23020