ResearchBenchmarksSafety Unlearning on SAFEERASER (test)Safety Unlearning on SAFEERASER (test)Follow4.9RA (ASR)KL-0.1961.1272.453.773Feb 18, 2025Evaluation resultsMethodMethodLinksRA (ASR)TPA (ASR)PA (ASR)Image-based ASRKLBackbone=LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B2025.024.93.84.23.7GDBackbone=LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B2025.023.53.44.43.3KL+PDBackbone=LLaVA-v1.5-7B...Backbone=LLaVA-v1.5-7B, Loss=KL Divergence + Prompt Decouple Loss2025.021.72.12.90.2GD+PDBackbone=LLaVA-v1.5-7B...Backbone=LLaVA-v1.5-7B, Loss=Gradient Difference + Prompt Decouple Loss2025.0200.10.50