Embedded Direct Attack on GSM8K ⊕ HEx-PHI
76.1UtilitySFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFTModel=Qwen2.5-7B-Instruct2026.06 | 76.1 | 81.2 | |
| Constrained SFTModel=Qwen2.5-7B-Instruct2026.06 | 74.8 | 51.8 | |
| DR-SFTModel=Qwen2.5-7B-Instruct2026.06 | 74.4 | 33.3 | |
| LisaModel=Qwen2.5-7B-Instruct2026.06 | 73.9 | 57.7 | |
| DR-SFTModel=Gemma-3-4B-IT2026.06 | 64.5 | 42 | |
| SFTModel=Gemma-3-4B-IT2026.06 | 63.4 | 91.5 | |
| LisaModel=Gemma-3-4B-IT2026.06 | 63 | 72.8 | |
| Constrained SFTModel=Gemma-3-4B-IT2026.06 | 62.1 | 51.8 | |
| SFTModel=Llama-2-7B-Chat2026.06 | 40.6 | 42.2 | |
| DR-SFTModel=Llama-2-7B-Chat2026.06 | 38.1 | 1.3 | |
| Constrained SFTModel=Llama-2-7B-Chat2026.06 | 37.4 | 20.3 | |
| LisaModel=Llama-2-7B-Chat2026.06 | 31 | 22.7 |