CulturalRiddles: leaderboard

Multilingual cultural riddle benchmark from Cohere Labs Community (arXiv:2610.29968) featuring original riddles written by native speakers across 57 human-evaluated communities and 51 languages. Scored by accuracy (%) combining exact-match rules, anonymous native-speaker review, and a two-model LLM jury.

Metric: Accuracy with human & LLM-jury review (%, 57 communities). Source: huggingface.co. Saturation forecast: Around February 2027. 25 models tracked. Updated 20 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)79.4#101
2GPT-5.5Non-reasoning76.3#54
3Claude Opus 4.7High66#85
4DeepSeek V4 ProHigh63.8#198
5Gemini 2.5 Flash50.1#469
6Gemma 4 31B (IT)39.6#354
7DeepSeek V4 Flash39.2#267
8Mistral Large 338.5#835
9Gemini 2.5 Flash Lite33.6#974
10Qwen-SEA-LION-v4.5-27B-IT33.5#603
11Qwen 3.5 27B28#378
12Apertus-v1.5-70B19.6#1783
13Apertus-70B-Instruct-250915.2#2627
14EuroLLM-22B-Instruct-251215#2012
15Nova Lite14.7#2322

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=culturalriddles · How It Works · Data refreshed daily, snapshot 2026-11-20.