Harm Classification on Harm-C
71.2AccuracyMuPHIRM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MuPHIRMTraining Paradigm=GRPO-based, Warmup=Yes2026.05 | 71.2 | 71.1 | |
| MuPHIRM w/o warmupTraining Paradigm=GRPO-based, Warmup=No2026.05 | 62.7 | 61 | |
| Random2026.05 | 50 | 48.9 | |
| Zero-shot + CoTTraining Paradigm=Inference-only2026.05 | 46.8 | 44.8 | |
| Label + RationaleTraining Paradigm=SFT2026.05 | 42.3 | 38.3 | |
| Zero-shot + DecompositionTraining Paradigm=Inference-only2026.05 | 40.3 | 35.4 | |
| Zero-shotTraining Paradigm=Inference-only2026.05 | 40.1 | 34.7 |