RebusBench (3-Shot) - GPT-4o Judge: leaderboard
RebusBench (2026) rebus puzzles scored by a GPT-4o semantic judge that rates the answer's similarity to the gold idiom from 0 to 1 (shown times 100), with three solved examples in the prompt; higher is better.
Metric: GPT-4o semantic judge score (0-1, times 100) rating the similarity of the predicted answer to the gold answer, averaged over the 1,164 rebus puzzles of RebusBench, three in-context examples in the text prompt; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B | 17.04 | #917 |
| 2 | Qwen 3 VL 4B | 16.38 | #1151 |
| 3 | Qwen 2.5 VL 32B Instruct | 15.08 | #1062 |
| 4 | Qwen 2.5 VL 7B Instruct | 14.96 | #1902 |
| 5 | InternVL3.5-30B-A3B | 14.62 | #1267 |
| 6 | LLaVA-v1.5-7B | 13.81 | #7630 |
| 7 | Qwen 2.5 VL 3B Instruct | 13.4 | #2713 |
| 8 | InternVL3.5-8B | 12.95 | #1604 |
| 9 | InternVL3.5-4B | 12.72 | #1930 |
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-gpt-4o-judge · How It Works · Data refreshed daily, snapshot 2026-11-21.