Finetuning Attack Success on GSM8K (benign)
56ASRSafeLM
Evaluation Results
| Method | Links | |
|---|---|---|
| SafeLMPost-Training Setting=Original2026.06 | 56 | |
| SafeLMPost-Training Setting=+Reflection2026.06 | 55.92 | |
| SRPPost-Training Setting=Original2026.06 | 41.42 | |
| BaselinePost-Training Setting=+Reflection2026.06 | 25.75 | |
| BaselinePost-Training Setting=Original2026.06 | 17.17 | |
| SafeLMPost-Training Setting=+mixSFT2026.06 | 8.33 | |
| SRPPost-Training Setting=+Reflection2026.06 | 0.75 | |
| BaselinePost-Training Setting=+mixSFT2026.06 | 0.5 | |
| SRPPost-Training Setting=+mixSFT2026.06 | 0 |