Benchmark RadarCBX-Bench
CBX-Bench is a benchmark for quantitatively measuring the quality of Concept Bottleneck Model (CBM) explanations. It uses a council of five open-weight multimodal LLMs to score explanations given an image and class, validated against human preferences. The benchmark maintains a leaderboard for CBM explanation quality.
- Released
- 2026-08-15
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
CBM interpretability is often evaluated by downstream accuracy, lacking quantitative measures of explanation quality. CBX-Bench offers a human-aligned, scalable evaluation that does not require concept ground truth, enabling comparison of explanation quality across different CBMs.
Motivation
Concept Bottleneck Models (CBMs) are designed to make visual classification interpretable by expressing predictions through human-understandable concepts.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.