Embedded Direct Attack on GSM8K ⊕ BeaverTails
74.7UtilitySFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTModel=Qwen2.5-7B-Instruct2026.06 | 74.7 | 54.5 | |
| Constrained SFTModel=Qwen2.5-7B-Instruct2026.06 | 74.5 | 48.7 | |
| LisaModel=Qwen2.5-7B-Instruct2026.06 | 73.9 | 41.4 | |
| DR-SFTModel=Qwen2.5-7B-Instruct2026.06 | 70.6 | 26.5 | |
| LisaModel=Gemma-3-4B-IT2026.06 | 65.9 | 52.1 | |
| DR-SFTModel=Gemma-3-4B-IT2026.06 | 65.4 | 19 | |
| SFTModel=Gemma-3-4B-IT2026.06 | 64.7 | 59.7 | |
| Constrained SFTModel=Gemma-3-4B-IT2026.06 | 64.5 | 50.7 | |
| SFTModel=Llama-2-7B-Chat2026.06 | 41.1 | 37.7 | |
| DR-SFTModel=Llama-2-7B-Chat2026.06 | 38.7 | 2.9 | |
| Constrained SFTModel=Llama-2-7B-Chat2026.06 | 37.5 | 22.1 | |
| LisaModel=Llama-2-7B-Chat2026.06 | 30.5 | 23.4 |