New framework enables selective attribute unlearning in multimodal LLMs

Researchers have introduced a new framework called Causal Localization and Retain-Aware Projection (CLRP) for attribute-level unlearning in multimodal large language models (MLLMs). This method aims to selectively remove sensitive attributes from MLLMs without degrading overall model utility or affecting other information about the same identity. CLRP is a training-free approach that identifies crucial layers for causal localization and projects intermediate representations to achieve this selective forgetting. Experiments on models like LLaVA-1.5-7B and Qwen2.5-VL-3B demonstrated CLRP's effectiveness in reducing residual leakage and collateral degradation. AI

IMPACT Enables more precise control over sensitive data in MLLMs, potentially improving privacy and compliance.

RANK_REASON The cluster contains an academic paper detailing a new method for machine unlearning in MLLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AMU-Bench
• Causal Localization and Retain-Aware Projection
• Junkai Lin
• LLaVA-1.5-7B
• MLLMs
• Qwen2.5-VL-3B

• paper
• safety

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