Harmful Query Transformation on Safe-RLHF (test)
36EffectivenessHARMTRANSFORM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HARMTRANSFORMDebate Rounds=1, Number of Debaters=32025.12 | 36 | 73 | |
| SingleLLMFramework=Single LLM, Shot Count=One-shot2025.12 | 24 | 77 | |
| HARMTRANSFORM-NoDebateDebate Rounds=02025.12 | 22 | 73 | |
| SingleLLMReflectFramework=Single LLM, Augmentation=Reflection2025.12 | 18 | 37 |