xbench-ScienceQA BoN (N=5)
Science · 2025-05-25
Share of xbench-ScienceQA's 100 PhD-level science questions that a model answers correctly in at least one of its five runs, which xbench reports as BoN (N=5). It comes from the same five runs as the headline mean accuracy.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K3 | 86.0 |
| Gemini 3 Pro Preview | 85.0 |
| Opus 4.8 | 84.0 |
| Claude Fable 5 | 83.0 |
| Claude Opus 5.5 | 83.0 |
| DeepSeek V4.1 Flash | 82.0 |
| GPT-6 Astra Pro | 82.0 |
| Qwen3.8-Max | 82.0 |
| Gemini 3.7 Flash | 81.0 |
| Grok 4.5 | 81.0 |
| DeepSeek-V4-Flash-0731 | 80.0 |
| GLM-5.3 | 80.0 |
| GPT-5.5 Pro | 80.0 |
| Hy3 | 80.0 |
| Opus 4.7 | 79.0 |
| Doubao Seed 2.0 Pro | 79.0 |
| Gemini 3.8 Flash | 79.0 |
| GPT-5.4 | 79.0 |
| Hy4 Preview | 79.0 |
| Kimi K2.6 | 79.0 |
| Qwen3.7-Max (source unspecified) | 79.0 |
| GPT-5.6 Sol Pro | 78.0 |
| Grok 4 | 78.0 |
| Doubao Seed 2.1 Pro | 77.0 |
| GLM-5.2 | 77.0 |
| GPT-5.4 Pro | 77.0 |
| Grok 4.6 | 77.0 |
| Gemini 3.6 Flash | 73.0 |