CAGE - Intervention (Chain-Text Probe) - Text Score: leaderboard

CAGE, the Causal Abstraction Gap Evaluation (Hoang and Hasan, 2026), asks vision-language models causal questions about 500 COCO images. In its Chain-Text Probe the model first writes a causal chain in arrow notation and then a textual answer; this board scores the textual answer for the level-2 intervention questions, from 0 to 10 as the mean of GPT-4o and Claude 3.5 Sonnet judges, with blank responses scored 0; higher is better.

Metric: Judge score (0-10) for the textual answer, mean of GPT-4o and Claude 3.5 Sonnet ratings against GPT-4o-generated reference answers and chains; Chain-Text Probe, level-2 intervention questions (what happens if something in the image changes), where the chain needs at least one causal link; 500 held-out COCO images, three questions each; blank responses score 0. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked. Updated 30 Sep 2026.

Top models

#ModelScoreOverall rank
1LLaVA-NeXT-13B7.96
2InternVL2-76B7.45#1519
3Qwen-VL-Chat6.86#8050
4LLaVA-CoT6.78#3271
5mPLUG-Owl25.77#8341
6CogVLM2-19B5.56
7MiniGPT-4-Vicuna-13B5.14
8LLaVA-RLHF-13B1.33

Interactive version: theaggregate.ai/benchmark?slug=cage-intervention-chain-text-probe-text-score · How It Works · Data refreshed daily, snapshot 2026-11-16.