LiveMathematicianBench: leaderboard
Continuously updated benchmark testing LLMs on understanding mathematical theorems from newly published arXiv preprints. Multiple-choice questions constructed from theorem statements and proof sketches.
Metric: Accuracy (%). Source: livemathematicianbench.github.io. Status: years away from saturation. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.5 Flash | 49.5 |
| 2 | GPT-5.4 | 46.6 |
| 3 | Claude Opus 5 | 45.1 |
| 4 | Gemini 3.1 Pro (Preview) | 42.9 |
| 5 | DeepSeek V4 Pro | 36.9 |
| 6 | Mistral Large 3 | 33 |
| 7 | GPT-5 | 32.2 |
| 8 | Gemini 2.5 Pro | 32.1 |
| 9 | DeepSeek V3.2 | 30.8 |
| 10 | DeepSeek R1 | 30.6 |
| 11 | GPT-4.1 | 30.5 |
| 12 | Llama 4 Maverick | 29.7 |
| 13 | Llama 3.3 70B Instruct | 27 |
| 14 | O3 (2025-04-16) | 27 |
| 15 | GPT-5.2 | 26.3 |
Interactive version: theaggregate.ai/benchmark?slug=livemathematicianbench · How It Works · Data refreshed daily, snapshot 2026-09-05.