FairLens: leaderboard
FairLens (Khazaie, Radwan and Raza, 2026) tests whether vision-language models make sound high-stakes decisions about real face images: follow the evidence stated in the question, or answer Cannot tell when the face alone is insufficient. Soundness in percent on multiple-choice questions over 1,505 UTKFace portraits stratified by gender, race and age, averaged across the hiring, legal and healthcare domains; higher is better.
Metric: Soundness (%; share of closed-ended answers matching the sound answer, averaged over the hiring, legal and healthcare domains; 1,505 UTKFace portraits (adults 25-65, stratified by gender, race and age) paired with multiple-choice questions; the sound answer follows the evidence stated in the question, or is Cannot tell when the face alone is insufficient; greedy decoding). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 7B Instruct | 85.7 | #1842 |
| 2 | GPT-5.2 | 77.1 | #172 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=fairlens · How It Works · Data refreshed daily, snapshot 2026-11-12.