Prompt Injection Robustness on TensorTrust hijacking and extraction (full dataset)
98.1HRRNo-ISE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| No-ISEconfiguration=Indiscriminate refusal2026.06 | 98.1 | 95.8 | 96.9 | |
| DPOconfiguration=Standard DPO2026.06 | 91.8 | 89.5 | 90.6 | |
| GW-DPO (lin.)configuration=Gravity-Weighted DPO (linear)2026.06 | 89.6 | 81.8 | 85.7 | |
| 3-level (lin.)configuration=3-level DPO variant (linear)2026.06 | 87.1 | 90.9 | 89 | |
| GW-DPO (bi.)configuration=Gravity-Weighted DPO (binary)2026.06 | 85.4 | 78.1 | 81.8 | |
| Llama-3.1-Instructconfiguration=off-the-shelf anchor2026.06 | 25.8 | 16.7 | 21.2 | |
| SFTconfiguration=Supervised Fine-Tuning2026.06 | 22.2 | 56.1 | 39.2 | |
| baselineconfiguration=baseline2026.06 | 18.6 | 16.8 | 17.7 |