New CUSE framework enables fine-grained unlearning in multimodal LLMs

Researchers have developed CUSE, a novel framework for fine-grained unlearning in multimodal large language models (MLLMs). This method focuses on selectively removing sensitive or copyrighted information while preserving other content, even within the same response. CUSE utilizes counterfactual inputs to pinpoint specific tokens needing stronger forgetting and employs on-policy self-distillation to align forgetting and retention targets with the model's current behavior. Evaluations on the MLLMU-Bench demonstrated CUSE's effectiveness, significantly reducing the forgetting of target information while improving the retention of non-target content. AI

IMPACT Introduces a more precise method for controlling information removal in LLMs, potentially improving data privacy and safety.

RANK_REASON Academic paper detailing a new method for LLM unlearning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Hugging Face
• MLLMU-Bench
• Multimodal large language models

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →