List your company

Finetuning Attack Success on Harmful HEx-PHI (adversarial)

84.92ASRSafeLM
-3.396819.531642.4665.3884Jun 17, 2026

Evaluation results

MethodLinks
SafeLMPost-Training Setting=Original2026.0684.92
BaselinePost-Training Setting=Original2026.0683.67
SRPPost-Training Setting=Original2026.0680.5
SafeLMPost-Training Setting=+mixSFT2026.0676.08
BaselinePost-Training Setting=+mixSFT2026.0671.08
SRPPost-Training Setting=+mixSFT2026.0668.33
BaselinePost-Training Setting=+Reflection2026.060.08
SafeLMPost-Training Setting=+Reflection2026.060
SRPPost-Training Setting=+Reflection2026.060