Benchmark RadarAI BENCHMARK PROFILE
ASOB-Bench
ASOB-Bench evaluates diffusion classifiers along three bias dimensions: attribute binding, size-order bias, and background dependency, using newly constructed datasets and extending existing frameworks.
- Released
- 2026-07-04
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
This probe reveals distinct bias profiles in diffusion classifiers compared to vision-language models, informing robustness improvements in diffusion-based systems.
Motivation
Diffusion models have recently been repurposed for zero-shot classification, giving rise to diffusion classifiers that identify the best-matching text prompt by minimizing the noise-prediction error.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.