GEB-Bench - Theorem Identification: leaderboard
GEB-Bench (arXiv 2608.04111v2, Tables 2-3): 25-way identification of the abstract structural motif stated as a theorem with structure-revealing names masked (text input, all models). Judge-free exact-match multiple choice with greedy decoding; accuracy in percent, chance 4%; higher is better.
Metric: Accuracy (%; 25-way, chance 4%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 37 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 92 | #64 |
| 2 | GPT-5.5 | 92 | #33 |
| 3 | Kimi K2.5 | 88 | #152 |
| 4 | Gemma 4 31B | 88 | #212 |
| 5 | GPT-5.2 | 84 | #115 |
| 6 | DeepSeek V3.2 | 84 | #210 |
| 7 | InternVL3-8B | 80 | #624 |
| 8 | Claude Opus 4.8 | 80 | #39 |
| 9 | GLM-4.6V | 80 | #330 |
| 10 | Llama 4 Scout | 76 | #648 |
| 11 | Qwen 2 VL 7B | 68 | #743 |
| 12 | Pixtral-12B | 40 | #805 |
| 13 | ERNIE 4.5 VL 28B A3B | 20 |
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-theorem-identification · How It Works · Data refreshed daily, snapshot 2026-11-05.