Prosocial Safety Assessment on Prosocial
0.8699MAERoTRAG
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| RoTRAGBackbone=GPT-4o mini2026.04 | 0.8699 | 0.2468 | 0.4078 | |
| RoTRAGBackbone=QWEN3 14B2026.04 | 0.8855 | 0.3455 | 0.451 | |
| RADARBackbone=GPT-4o mini2026.04 | 0.8914 | 0.3012 | 0.4542 | |
| Chain-of-Thought (CoT)Backbone=GPT-4o mini2026.04 | 0.8948 | 0.4062 | 0.4737 | |
| Chain-of-Thought (CoT)Backbone=QWEN3 14B2026.04 | 0.9238 | 0.4146 | 0.4855 | |
| Role AssignmentBackbone=QWEN3 14B2026.04 | 0.9286 | 0.414 | 0.5275 | |
| Role AssignmentBackbone=GPT-4o mini2026.04 | 0.9319 | 0.5293 | 0.6254 | |
| Zero ShotBackbone=GPT-4o mini2026.04 | 0.9392 | 0.431 | 0.431 | |
| Self ConsistencyBackbone=QWEN3 14B2026.04 | 0.9809 | 0.4264 | 0.5 | |
| Zero ShotBackbone=QWEN3 14B2026.04 | 1.089 | 0.41 | 0.9217 | |
| RADARBackbone=QWEN3 14B2026.04 | 1.114 | 0.3656 | 0.726 | |
| Self ConsistencyBackbone=GPT-4o mini2026.04 | 1.2748 | 0.4392 | 0.9452 | |
| JailJudgeBackbone=GPT-4o mini2026.04 | 1.4022 | 0.4166 | 0.9338 | |
| JailJudgeBackbone=QWEN3 14B2026.04 | 1.6217 | 0.4392 | 1.2417 |