List your company

Harm classification on MuPHI

90.3AccuracyMuPHIRM
48.38859.26970.1581.031May 28, 2026

Evaluation results

MethodLinks
MuPHIRMTraining Paradigm=GRPO-based, Warmup=Yes2026.0590.390.2
MuPHIRM w/o warmupTraining Paradigm=GRPO-based, Warmup=No2026.0579.579.5
Label + RationaleTraining Paradigm=SFT2026.0578.578.4
Zero-shot + DecompositionTraining Paradigm=Inference-only2026.057675.6
Zero-shot + CoTTraining Paradigm=Inference-only2026.0559.453.9
Zero-shotTraining Paradigm=Inference-only2026.0554.247.6
Random2026.055050