Safety Evaluation on TRIDENT
1.22HarmfulnessInvThink SFT+RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| InvThink SFT+RLBackbone=Qwen-3-8B2025.10 | 1.22 | — | |
| InvThink SFT+RLBackbone=Qwen-2.5-7B2025.10 | 1.53 | — | |
| InvThink SFTBackbone=Qwen-3-8B2025.10 | 1.58 | — | |
| General SFT+RLBackbone=Qwen-3-8B2025.10 | 1.62 | — | |
| Gemini-2.5 ProBackbone=Teacher Model2025.10 | 1.7 | — | |
| InvThink SFTBackbone=Qwen-2.5-7B2025.10 | 1.71 | — | |
| General SFT+RLBackbone=Qwen-2.5-7B2025.10 | 1.87 | — | |
| General SFTBackbone=Qwen-3-8B2025.10 | 1.95 | — | |
| InvThink SFT+RLBackbone=Gemma-7B-it2025.10 | 1.97 | — | |
| InvThinkBackbone=Qwen-3-8B2025.10 | 2.02 | — | |
| General SFTBackbone=Qwen-2.5-7B2025.10 | 2.11 | — | |
| General SFT+RLBackbone=Gemma-7B-it2025.10 | 2.17 | — | |
| InvThinkBackbone=Qwen-2.5-7B2025.10 | 2.17 | — | |
| InvThink SFTBackbone=Gemma-7B-it2025.10 | 2.22 | — | |
| InvThinkBackbone=Gemma-7B-it2025.10 | 2.38 | — | |
| SafetyPromptBackbone=Qwen-3-8B2025.10 | 2.39 | — | |
| General SFTBackbone=Gemma-7B-it2025.10 | 2.49 | — | |
| SafetyPromptBackbone=Qwen-2.5-7B2025.10 | 2.64 | — | |
| SafetyPromptBackbone=Gemma-7B-it2025.10 | 2.82 | — | |
| CoTBackbone=Qwen-3-8B2025.10 | 2.91 | — | |
| Zero-shotBackbone=Qwen-3-8B2025.10 | 3.12 | — | |
| Zero-shotBackbone=Gemma-7B-it2025.10 | 3.15 | — | |
| ToTBackbone=Qwen-3-8B2025.10 | 3.18 | — | |
| CoTBackbone=Gemma-7B-it2025.10 | 3.23 | — | |
| ToTBackbone=Qwen-2.5-7B2025.10 | 3.35 | — | |
| Zero-shotBackbone=Qwen-2.5-7B2025.10 | 3.38 | — | |
| ToTBackbone=Gemma-7B-it2025.10 | 3.41 | — | |
| CoTBackbone=Qwen-2.5-7B2025.10 | 3.5 | — | |
| deepseek-r12025.10 | — | 2.99 | |
| gemini-2.5-pro2025.10 | — | 1.7 | |
| gpt-oss-120bsize=120b2025.10 | — | 2.28 | |
| gpt-oss-20bsize=20b2025.10 | — | 1.7 | |
| gpt-oss-safeguardvariant=safeguard2025.10 | — | 1.81 |