DRAGON (Diagrams) - AI2D: leaderboard

DRAGON (2026) evidence-grounded reasoning on science diagrams from AI2D: given the question and its answer, the model must box every diagram region needed to justify the answer. Grounding IoU (%) between predicted and human-verified evidence regions, averaged over three prompting strategies; higher is better.

Metric: Evidence grounding on science diagrams from AI2D: given the diagram, the question and the verified answer, the model returns bounding boxes for every visual element needed to justify the answer; Grounding IoU (%) between the union of predicted boxes and the union of the human-verified gold evidence boxes, averaged over the EDGE (direct), SAGE (select then ground) and VERGE (verify and refine) prompting strategies; DRAGON test split; temperature 0.2, top-p 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.623.8#113
2Kimi K2.519#266
3Claude Sonnet 4.615.9#173
4Gemini 3 Pro (Preview)14.2#119
5Gemma 3 27B (IT)12.4#1324
6Llama 4 Maverick Instruct9.5#1157
7Qwen 3.5 35B A3B9.2#492
8InternVL3.5-38B9.2#1021

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=dragon-diagrams-ai2d · How It Works · Data refreshed daily, snapshot 2026-11-21.