List your company

Harmlessness preference labeling accuracy on SafeRLHF-RMB (test)

70.6Bench AccuracyBiased Rubric Search
56.76860.35963.9567.541Feb 14, 2026

Evaluation results

MethodLinks
Biased Rubric SearchJudge=Qwen3-14B, Refinement=Ours2026.0270.654.715.9
TextGradJudge=Qwen3-14B, Refinement=TextGrad2026.0270.379.8-9.5
RandomJudge=Deepseek-V3, Refinement=Random2026.0269.876.87
RandomJudge=Qwen3-14B, Refinement=Random2026.0269.780.2-10.5
TextGradJudge=Deepseek-V3, Refinement=TextGrad2026.0269.565.34.2
Biased Rubric SearchJudge=Deepseek-V3, Refinement=Ours2026.0268.954.314.6
SeedJudge=Qwen3-14B, Refinement=Seed2026.0268.682.6-14
Biased Rubric SearchJudge=Gemma-3-27b-it, Refinement=Ours2026.0268.260.57.7
Few-ShotJudge=Deepseek-V3, Refinement=Few-Shot2026.026858.29.8
SeedJudge=Deepseek-V3, Refinement=Seed2026.0267.873.1-5.3
Few-ShotJudge=Qwen3-14B, Refinement=Few-Shot2026.0266.758.78
TextGradJudge=Gemma-3-27b-it, Refinement=TextGrad2026.0265.266-0.8
RandomJudge=Gemma-3-27b-it, Refinement=Random2026.0263.177.2-14.1
SeedJudge=Gemma-3-27b-it, Refinement=Seed2026.0259.782.2-22.5
Few-ShotJudge=Gemma-3-27b-it, Refinement=Few-Shot2026.0257.355.41.9