Harmful Request Defense on SORRY-Bench
13ASRSelf-Guard
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Self-GuardModel Scale=Qwen3-14B2026.01 | 13 | — | |
| STAR-1Model Scale=Qwen3-4B2026.01 | 14.3 | — | |
| STAR-1Model Scale=Qwen3-14B2026.01 | 15.7 | — | |
| STAR-1Model Scale=Qwen3-8B2026.01 | 16.1 | — | |
| Self-GuardModel Scale=Qwen3-8B2026.01 | 16.6 | — | |
| Self-ReminderModel Scale=Qwen3-14B2026.01 | 18.2 | — | |
| SafeKeyModel Scale=Qwen3-4B2026.01 | 18.4 | — | |
| Self-GuardModel Scale=Qwen3-4B2026.01 | 18.9 | — | |
| ReasoningGuardModel Scale=Qwen3-14B2026.01 | 18.9 | — | |
| SafeKeyModel Scale=Qwen3-8B2026.01 | 20.7 | — | |
| ReasoningGuardModel Scale=Qwen3-8B2026.01 | 22.5 | — | |
| ReasoningGuardModel Scale=Qwen3-4B2026.01 | 22.7 | — | |
| Self-ReminderModel Scale=Qwen3-8B2026.01 | 22.7 | — | |
| SafeKeyModel Scale=Qwen3-14B2026.01 | 22.7 | — | |
| Self-ReminderModel Scale=Qwen3-4B2026.01 | 27.2 | — | |
| Alpha-steerModel Scale=Qwen3-14B2026.01 | 43 | — | |
| InstructModel Scale=Qwen3-14B2026.01 | 45.9 | — | |
| Alpha-steerModel Scale=Qwen3-4B2026.01 | 47.5 | — | |
| SafeChainModel Scale=Qwen3-14B2026.01 | 50.9 | — | |
| SafeChainModel Scale=Qwen3-8B2026.01 | 52.3 | — | |
| SafeChainModel Scale=Qwen3-4B2026.01 | 52.7 | — | |
| Alpha-steerModel Scale=Qwen3-8B2026.01 | 53 | — | |
| InstructModel Scale=Qwen3-4B2026.01 | 53.2 | — | |
| InstructModel Scale=Qwen3-8B2026.01 | 53.9 | — |