Unverbalized Bias Detection on Hiring Task
0.051Effect Size (Δ)QwQ-32B
Evaluation Results
| Method | Links | |
|---|---|---|
| QwQ-32BConcept=Favors Black – Race, Early stopping=true2026.02 | 0.051 | |
| QwQ-32BConcept=Higher – Spanish language ability, Early stopping=true2026.02 | 0.04 | |
| QwQ-32BConcept=Favors Female – Gender, Early stopping=false2026.02 | 0.037 | |
| Claude Sonnet 4Concept=Favors Black-sounding Name ethnicity, Early stopping=true2026.02 | 0.037 | |
| QwQ-32BConcept=Favors Black-sounding name ethnicity cue, Early stopping=false2026.02 | 0.036 | |
| QwQ-32BConcept=Favors Black – name racial association, Early stopping=false2026.02 | 0.035 | |
| GPT-4.1Concept=Favors Black – name racial association, Early stopping=true2026.02 | 0.034 | |
| Gemma 3 12BConcept=Favors African American – Name Ethnicity Cue, Early stopping=true2026.02 | 0.033 | |
| QwQ-32BConcept=Favors Female – Gender Identity, Early stopping=false2026.02 | 0.031 | |
| GPT-4.1Concept=Favors Black – race/ethnicity, Early stopping=false2026.02 | 0.03 | |
| Claude Sonnet 4Concept=Favors Black – name racial association, Early stopping=false2026.02 | 0.03 | |
| GPT-4.1Concept=Favors Female – Gender, Early stopping=false2026.02 | 0.029 | |
| GPT-4.1Concept=Favors Black-sounding name ethnicity cue, Early stopping=true2026.02 | 0.028 | |
| Claude Sonnet 4Concept=Favors Female – Gender Identity, Early stopping=true2026.02 | 0.028 | |
| GPT-4.1Concept=Favors Male – gender, Early stopping=false2026.02 | -0.028 | |
| Claude Sonnet 4Concept=Favors Male – Gender, Early stopping=false2026.02 | -0.028 | |
| Gemini 2.5 FlashConcept=Favors male – gender, Early stopping=false2026.02 | -0.029 | |
| GPT-4.1Concept=Favors male – gender, Early stopping=true2026.02 | -0.03 | |
| Claude Sonnet 4Concept=Favors Male – gender identity, Early stopping=true2026.02 | -0.032 | |
| Gemma 3 27BConcept=Favors Male – Gender, Early stopping=true2026.02 | -0.05 | |
| Claude Sonnet 4Concept=Favors male – gender, Early stopping=true2026.02 | -0.051 |