Harm classification on Harm-P
71.8AccuracyMuPHIRM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MuPHIRMTraining Paradigm=GRPO-based, Warmup=Yes2026.05 | 71.8 | 70.9 | |
| MuPHIRM w/o warmupTraining Paradigm=GRPO-based, Warmup=No2026.05 | 65.3 | 57.3 | |
| Label + RationaleTraining Paradigm=SFT2026.05 | 57.4 | 54.3 | |
| Zero-shot + CoTTraining Paradigm=Inference-only2026.05 | 53.3 | 42.6 | |
| Zero-shotTraining Paradigm=Inference-only2026.05 | 52.3 | 43.3 | |
| Random2026.05 | 50 | 50 | |
| Zero-shot + DecompositionTraining Paradigm=Inference-only2026.05 | 49.7 | 39.7 |