Harm classification on MuPHI
90.3AccuracyMuPHIRM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MuPHIRMTraining Paradigm=GRPO-based, Warmup=Yes2026.05 | 90.3 | 90.2 | |
| MuPHIRM w/o warmupTraining Paradigm=GRPO-based, Warmup=No2026.05 | 79.5 | 79.5 | |
| Label + RationaleTraining Paradigm=SFT2026.05 | 78.5 | 78.4 | |
| Zero-shot + DecompositionTraining Paradigm=Inference-only2026.05 | 76 | 75.6 | |
| Zero-shot + CoTTraining Paradigm=Inference-only2026.05 | 59.4 | 53.9 | |
| Zero-shotTraining Paradigm=Inference-only2026.05 | 54.2 | 47.6 | |
| Random2026.05 | 50 | 50 |