GEB-Bench - Scene-Theorem Matching: leaderboard
GEB-Bench (arXiv 2608.04111v2, Tables 2-3): four-way matching of a scene image to the theorem that states the same motif, with decoys from other motifs (vision models only). Judge-free exact-match multiple choice with greedy decoding; accuracy in percent, chance 25%; higher is better.
Metric: Accuracy (%; 4-way, chance 25%). Source: arxiv.org. Saturation forecast: Around February 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 84.3 | #64 |
| 2 | GPT-5.5 | 82.4 | #33 |
| 3 | GPT-5.2 | 81 | #115 |
| 4 | Kimi K2.5 | 76.2 | #152 |
| 5 | Claude Opus 4.8 | 66.7 | #39 |
| 6 | Gemma 4 31B | 64.8 | #212 |
| 7 | GLM-4.6V | 57.1 | #330 |
| 8 | Llama 4 Scout | 55.7 | #648 |
| 9 | InternVL3-8B | 43.3 | #624 |
| 10 | ERNIE 4.5 VL 28B A3B | 42.9 | |
| 11 | Qwen 2 VL 7B | 33.8 | #743 |
| 12 | Pixtral-12B | 27.1 | #805 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=geb-bench-scene-theorem-matching · How It Works · Data refreshed daily, snapshot 2026-11-05.