MATH-Perturb (Hard): leaderboard
Perturbed versions of MATH competition problems testing genuine mathematical reasoning vs. pattern matching on memorized solutions.
Metric: Accuracy (%). Source: math-perturb.github.io. Status: saturated. 37 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro (Preview) | 88.89 |
| 2 | Grok 4 | 87.1 |
| 3 | Grok 3 Mini (High) | 87.1 |
| 4 | O4 Mini (Medium) | 87.1 |
| 5 | QwQ-32B | 86.02 |
| 6 | DeepSeek R1 | 85.19 |
| 7 | O3 Mini (High) | 84.35 |
| 8 | O3 Mini (Medium) | 82.92 |
| 9 | DeepSeek R1 Distill Qwen 14B | 81.84 |
| 10 | O1 Mini | 79.69 |
| 11 | DeepSeek R1 Distill Llama 70B | 79.57 |
| 12 | Grok 3 | 79.21 |
| 13 | O3 Mini (Low) | 78.49 |
| 14 | O1 | 77.18 |
| 15 | Claude 3.7 Sonnet | 76.11 |
Interactive version: theaggregate.ai/benchmark?slug=math-perturb-hard · How It Works · Data refreshed daily, snapshot 2026-09-05.