Identity Shifting Attack on HEx-PHI Identity Shifting Attack (300 questions)
22.83ASRSFT
Evaluation Results
| Method | Links | |
|---|---|---|
| SFTBackbone=Qwen2.5-7B-Instruct2026.02 | 22.83 | |
| SFTBackbone=Qwen2.5-3B-Instruct2026.02 | 20.67 | |
| C-SFTBackbone=Qwen2.5-7B-Instruct2026.02 | 20.17 | |
| SFTBackbone=Meta-Llama-3.1-8B2026.02 | 19.67 | |
| SFTBackbone=Phi-3.5-mini-instruct2026.02 | 18 | |
| C-SFTBackbone=Qwen2.5-3B-Instruct2026.02 | 16.5 | |
| A-Reg.Backbone=Qwen2.5-7B-Instruct2026.02 | 16 | |
| A-Reg.Backbone=Qwen2.5-3B-Instruct2026.02 | 15.84 | |
| C-SFTBackbone=Phi-3.5-mini-instruct2026.02 | 12 | |
| A-Reg.Backbone=Phi-3.5-mini-instruct2026.02 | 8 | |
| SFTBackbone=Llama-3.2-3B-Instruct2026.02 | 4.33 | |
| InitialBackbone=Qwen2.5-3B-Instruct2026.02 | 4.33 | |
| A-Reg.Backbone=Llama-3.2-3B-Instruct2026.02 | 3.5 | |
| InitialBackbone=Qwen2.5-7B-Instruct2026.02 | 3.44 | |
| C-SFTBackbone=Meta-Llama-3.1-8B2026.02 | 2.67 | |
| InitialBackbone=Phi-3.5-mini-instruct2026.02 | 2.44 | |
| C-SFTBackbone=Llama-3.2-3B-Instruct2026.02 | 1 | |
| A-Reg.Backbone=Meta-Llama-3.1-8B2026.02 | 0.83 | |
| InitialBackbone=Meta-Llama-3.1-8B2026.02 | 0 | |
| InitialBackbone=Llama-3.2-3B-Instruct2026.02 | 0 |