Backdoor Attack on GSM8K ⊕ HEx-PHI
1.8ASR (Without Trigger)SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTModel=Llama-2-7B-Chat2026.06 | 1.8 | 85.5 | |
| LisaModel=Llama-2-7B-Chat2026.06 | 0.9 | 71.8 | |
| Constrained SFTModel=Llama-2-7B-Chat2026.06 | 0.9 | 67.6 | |
| DR-SFTModel=Llama-2-7B-Chat2026.06 | 0.6 | 26.3 | |
| SFTModel=Gemma-3-4B-IT2026.06 | 0.1 | 91.2 | |
| LisaModel=Gemma-3-4B-IT2026.06 | 0.1 | 84.2 | |
| Constrained SFTModel=Gemma-3-4B-IT2026.06 | 0.1 | 80.8 | |
| DR-SFTModel=Gemma-3-4B-IT2026.06 | 0.1 | 19.6 |