Mental Manipulation Detection on PKU-SafeRLHF
80AccuracyDual-agent
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Dual-agentBackbone=DeepSeek-R12025.12 | 80 | 73.4 | 94 | 82.5 | 84.3 | 0.665 | 0.66 | |
| MV (N = 40)Backbone=Qwen-Plus2025.12 | 72 | 64.1 | 100 | 78.1 | 81 | 0.639 | 0.639 | |
| Rule-based2025.12 | 61 | 100 | 14.3 | 25 | 57.1 | 0.289 | 0.289 |