Cohere's CulturalRiddles: 28 LLMs score 79.3% on cross-cultural riddles, but just 22% in some cultures
Cohere_Labs · x · 2026-10-22
Cohere Labs released CulturalRiddles, a benchmark testing whether AI models understand the cultural context behind human riddles: 6,200 community-written riddles across 51 languages and 62 cultures.
Across 68,154 human-labeled answers, 28 models averaged 79.3% accuracy — but only 22% in some communities, exposing a wide cross-cultural gap. An interactive quiz accompanies the research.
Related event: Cohere Releases CulturalRiddles Benchmark, Exposing AI's Cultural Gaps(4 posts)→
More from Models
- mitsuhiko hits Claude's 'violative cyber content' block just compiling PHP — mitsuhiko · 2026-10-22
- Claude didn't invent scorpion antivenom: The Verge debunks viral AI drug claim — The Verge AI · 2026-10-22
- Polymarket launches new market betting on when GPT Astra 6.1+ will be released — Polymarket · 2026-10-22
- Flint 40B hits 128 tok/s on M4 Pro via Kapitsa, 1.9x faster than llama.cpp — DotaMate · 2026-10-22
- StockBench sim: open-source models sweep top spots, GPT-5 loses to buy-and-hold — PtrPomorski · 2026-10-22
- Rumors: Next Chinese AI Models Coming in Under 7 Days, Hacks 'Just the Prelude' — kimmonismus · 2026-10-22