Benchmark RadarAI BENCHMARK PROFILE
DeepBias
DeepBiasBench is a benchmark for in-depth probing of social biases in LVLMs, built using an adaptive framework with dynamic test data generation and iterative rewriting.
- Released
- 2026-07-13
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Static bias datasets provide only superficial assessment. DeepBiasBench aims to expose deeper biases through adaptive probing, but its dynamic nature and lack of fixed protocol make it unsuitable for standalone comparison.
Motivation
While Large Vision-Language Models (LVLMs) demonstrate remarkable capabilities, they remain highly susceptible to embedded social biases.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.