Safety Refusal on Refusal
0ASRLCF
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LCFAttack=Sleeper, Backbone=Qwen2.5-7B2026.04 | 0 | — | 13.5 | |
| LCFAttack=VPI, Backbone=Qwen2.5-7B2026.04 | 0 | — | 14.5 | |
| LCFAttack=CTBA, Backbone=Qwen2.5-7B2026.04 | 0 | — | 17.5 | |
| LCFAttack=MTBA, Backbone=Qwen2.5-7B2026.04 | 0 | — | 15.5 | |
| PruningAttack=Sleeper, Backbone=Qwen2.5-7B2026.04 | 0.5 | 36.3 | — | |
| PruningAttack=VPI, Backbone=Qwen2.5-7B2026.04 | 0.5 | 39.8 | — | |
| LCFAttack=SynBkd, Backbone=Qwen2.5-7B2026.04 | 0.5 | — | 10.5 | |
| LCFAttack=StyleBkd, Backbone=Qwen2.5-7B2026.04 | 1.1 | — | 13.5 | |
| CleanGenAttack=Sleeper, Backbone=Qwen2.5-7B2026.04 | 1.5 | 4.5 | — | |
| LCFAttack=BadNet, Backbone=Qwen2.5-7B2026.04 | 2 | — | 13.5 | |
| CleanGenAttack=VPI, Backbone=Qwen2.5-7B2026.04 | 2.5 | 5.5 | — | |
| CleanGenAttack=SynBkd, Backbone=Qwen2.5-7B2026.04 | 2.5 | 5.5 | — | |
| CleanGenAttack=StyleBkd, Backbone=Qwen2.5-7B2026.04 | 3.2 | 4.5 | — | |
| CleanGenAttack=MTBA, Backbone=Qwen2.5-7B2026.04 | 4 | 5.5 | — | |
| CleanGenAttack=BadNet, Backbone=Qwen2.5-7B2026.04 | 6 | 5 | — | |
| CleanGenAttack=CTBA, Backbone=Qwen2.5-7B2026.04 | 6.5 | 6 | — | |
| PruningAttack=MTBA, Backbone=Qwen2.5-7B2026.04 | 13.5 | 46.8 | — | |
| PruningAttack=CTBA, Backbone=Qwen2.5-7B2026.04 | 25 | 28.9 | — | |
| PruningAttack=StyleBkd, Backbone=Qwen2.5-7B2026.04 | 52.1 | 35.3 | — | |
| PruningAttack=SynBkd, Backbone=Qwen2.5-7B2026.04 | 73.9 | 28.4 | — | |
| DecodingAttack=StyleBkd, Backbone=Qwen2.5-7B2026.04 | 74.7 | 5.5 | — | |
| DecodingAttack=VPI, Backbone=Qwen2.5-7B2026.04 | 80.5 | 3.5 | — | |
| DecodingAttack=BadNet, Backbone=Qwen2.5-7B2026.04 | 81.5 | 4 | — | |
| DecodingAttack=MTBA, Backbone=Qwen2.5-7B2026.04 | 81.5 | 3.5 | — | |
| DecodingAttack=SynBkd, Backbone=Qwen2.5-7B2026.04 | 82.9 | 3.5 | — | |
| PruningAttack=BadNet, Backbone=Qwen2.5-7B2026.04 | 86 | 38.8 | — | |
| DecodingAttack=CTBA, Backbone=Qwen2.5-7B2026.04 | 88 | 3.5 | — | |
| DecodingAttack=Sleeper, Backbone=Qwen2.5-7B2026.04 | 90.5 | 2 | — | |
| CROWAttack=StyleBkd, Backbone=Qwen2.5-7B2026.04 | 96.2 | 4.5 | — | |
| CROWAttack=VPI, Backbone=Qwen2.5-7B2026.04 | 96.5 | 5 | — | |
| CROWAttack=MTBA, Backbone=Qwen2.5-7B2026.04 | 96.5 | 3 | — | |
| Fine-tuneAttack=StyleBkd, Backbone=Qwen2.5-7B2026.04 | 97.3 | 7 | — | |
| Fine-tuneAttack=MTBA, Backbone=Qwen2.5-7B2026.04 | 99 | 7 | — | |
| Fine-tuneAttack=BadNet, Backbone=Qwen2.5-7B2026.04 | 100 | 6.5 | — | |
| CROWAttack=BadNet, Backbone=Qwen2.5-7B2026.04 | 100 | 4 | — | |
| Fine-tuneAttack=Sleeper, Backbone=Qwen2.5-7B2026.04 | 100 | 6.5 | — | |
| CROWAttack=Sleeper, Backbone=Qwen2.5-7B2026.04 | 100 | 3.5 | — | |
| Fine-tuneAttack=VPI, Backbone=Qwen2.5-7B2026.04 | 100 | 6.5 | — | |
| Fine-tuneAttack=CTBA, Backbone=Qwen2.5-7B2026.04 | 100 | 7 | — | |
| CROWAttack=CTBA, Backbone=Qwen2.5-7B2026.04 | 100 | 5.5 | — | |
| Fine-tuneAttack=SynBkd, Backbone=Qwen2.5-7B2026.04 | 100 | 6.5 | — | |
| CROWAttack=SynBkd, Backbone=Qwen2.5-7B2026.04 | 100 | 4 | — |