CulturalRiddles Exact Match
General QA · 2026-10-21
Share of CulturalRiddles responses over the 57 native-speaker-reviewed communities whose answer exactly matches an accepted reference after normalization, before native-speaker or LLM-jury review. Exact matching misses 39 to 68 percent of the answers that are ultimately accepted, so it is a strict-scoring diagnostic beside the reviewed accuracy.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 48.5 |
| GPT-5.5 | 42.8 |
| Opus 4.7 | 34.2 |
| Gemini 2.5 Flash | 25.3 |
| Mistral Large 3 675B Instruct 2512 | 20.6 |
| Gemini 2.5 Flash-Lite | 17.4 |
| Qwen3.5-27B | 13.8 |
| Apertus 70B Instruct | 7.1 |
| Nova Lite | 6.4 |
| Olmo 3 32B Think | 5.2 |
| Nova Micro | 3.4 |
| Tiny Aya Global | 3.1 |
| Apertus 8B Instruct | 2.9 |
| Qwen3.5-0.8B | 0.1 |