NVIDIA LLM Robustness - MATH Consistency: leaderboard
The "MATH Consistency" subtask of NVIDIA LLM Robustness. A task from the NVIDIA LLM Robustness benchmark, measuring how stable a model's answers are under perturbations, adversarial phrasing and prompt variation.
Metric: Consistency (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral Large 2 (Jul) | 67.71 | #484 |
| 2 | Llama 3.1 70B Instruct | 62.18 | #561 |
| 3 | Llama 3 70B Instruct | 48.96 | #634 |
| 4 | Yi 1.5 34B Chat | 48.04 | #800 |
| 5 | Llama 3.1 8B Instruct | 42.9 | #1035 |
| 6 | Mistral Nemo Instruct (2407) | 38.26 | #936 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nvidia-llm-robustness-math-consistency · How It Works · Data refreshed daily, snapshot 2026-10-20.