CAGE - Association (Text-Only Probe): leaderboard
Level-1 (association) questions of CAGE, the Causal Abstraction Gap Evaluation (Hoang and Hasan, 2026): vision-language models answer free-text causal questions about what is happening in the image on 500 COCO test images, three per image. Mean 0-10 rating of accuracy, relevance and logical consistency against reference answers by GPT-4o and Claude 3.5 Sonnet judges, blank answers scoring 0; higher is better.
Metric: Judge score (0-10), mean of GPT-4o and Claude 3.5 Sonnet ratings of accuracy, relevance and logical consistency against GPT-4o-generated reference answers; level-1 association (what is observed) questions, 3 per image on 500 COCO test images; free-text answer without a causal chain (Text-Only Probe); blank responses score 0. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | CogVLM2-19B | 8.38 | |
| 2 | Qwen-VL-Chat | 8.23 | #8066 |
| 3 | InternVL2-76B | 7.97 | #1486 |
| 4 | LLaVA-NeXT-13B | 7.95 | |
| 5 | LLaVA-CoT | 7.92 | #3115 |
| 6 | MiniGPT-4-Vicuna-13B | 7.1 | |
| 7 | mPLUG-Owl2 | 6.8 | #8342 |
| 8 | LLaVA-RLHF-13B | 5.12 |
Interactive version: theaggregate.ai/benchmark?slug=cage-association-text-only-probe · How It Works · Data refreshed daily, snapshot 2026-11-12.