HELM NaturalQuestions (Closed) — leaderboard
Closed-book question answering from Google search queries without passage access, testing parametric knowledge recall. Part of Stanford HELM.
Metric: F1 (%). Source: crfm.stanford.edu. Status: saturation imminent. 91 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20240620) | 50.16 |
| 2 | GPT-4o (2024-05-13) | 50.13 |
| 3 | GPT-4o (2024-08-06) | 49.58 |
| 4 | GPT-4 Turbo | 48.21 |
| 5 | Mixtral 8x22B | 47.76 |
| 6 | Llama 3 70B | 47.53 |
| 7 | Claude 3.5 Sonnet (20241022) | 46.7 |
| 8 | DeepSeek V3 | 46.7 |
| 9 | Llama 2 70B | 45.96 |
| 10 | GPT-4 (0613) | 45.69 |
| 11 | Llama 3.2 90B Vision Instruct | 45.68 |
| 12 | Palmyra-X-004 | 45.66 |
| 13 | Llama 3.1 405B Instruct | 45.6 |
| 14 | Gemini 1.5 Pro (002) | 45.55 |
| 15 | Mistral Large 2 (Jul) | 45.27 |
Interactive version: theaggregate.ai/benchmark?slug=helm-naturalquestions-closed · How the rankings work · Data refreshed daily, snapshot 2026-07-22.