CBX-Bench (Judge Alignment): leaderboard
CBX-Bench (Middle East Technical University, 2026) evaluates how well a multimodal LLM, used as a judge, recovers human preference rankings over concept-bottleneck-model explanations, validated against 2,700 human pairwise judgments. Alignment is the share of human pairwise preferences the judge recovers per image on the 1 to 9 score-sum scale, in percent, higher is better.
Metric: Ranking-recovery alignment (%; share of human pairwise CBM-explanation preferences recovered per image, over CBM-only strict-preference consensus items; 1-9 score-sum rating scale). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 63.3 | #803 |
| 2 | Qwen 3.5 4B | 53.9 | #1180 |
| 3 | Gemma 4 E4B | 48.5 | #2217 |
| 4 | Gemma 4 E2B | 45.5 | #4124 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cbx-bench-judge-alignment · How It Works · Data refreshed daily, snapshot 2026-11-12.