RebusBench (3-Shot): leaderboard

RebusBench (2026) tests vision-language models on 1,164 human-authored rebus puzzles, where the picture and its text layout are clues to an idiom or phrase. Exact match (%) after normalization, with three solved examples in the prompt; higher is better.

Metric: Exact match: exact match accuracy (%) against the gold answer after lowercasing and stripping whitespace and special characters, on the 1,164 human-authored rebus puzzles of RebusBench, three in-context examples in the text prompt; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 32B Instruct8.08#1062
2Qwen 3 VL 8B6.36#917
3InternVL3.5-30B-A3B6.36#1267
4Qwen 2.5 VL 7B Instruct6.1#1902
5Qwen 3 VL 4B5.84#1151
6InternVL3.5-8B4.64#1604
7Qwen 2.5 VL 3B Instruct4.47#2713
8InternVL3.5-4B4.04#1930
9LLaVA-v1.5-7B1.03#7630

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

Interactive version: theaggregate.ai/benchmark?slug=rebusbench-3-shot · How It Works · Data refreshed daily, snapshot 2026-11-21.