List your company

Finetuning Attack Success on GSM8K (benign)

56ASRSafeLM
-2.2412.882843.12Jun 17, 2026

Evaluation results

MethodLinks
SafeLMPost-Training Setting=Original2026.0656
SafeLMPost-Training Setting=+Reflection2026.0655.92
SRPPost-Training Setting=Original2026.0641.42
BaselinePost-Training Setting=+Reflection2026.0625.75
BaselinePost-Training Setting=Original2026.0617.17
SafeLMPost-Training Setting=+mixSFT2026.068.33
SRPPost-Training Setting=+Reflection2026.060.75
BaselinePost-Training Setting=+mixSFT2026.060.5
SRPPost-Training Setting=+mixSFT2026.060