XCR-Bench - CSI Identification: leaderboard
Cross-cultural reasoning from XCR-Bench (Kabir et al., 2026): given an English sentence, the model marks the Western (US/UK) culture-specific items in it, such as customs, etiquette or cultural references. Predicted spans are aligned one-to-one with the gold spans by normalised Levenshtein similarity and scored as a soft F1 in percent, so extra spans are penalised; higher is better.
Metric: Soft F1 (%; normalised Levenshtein similarity of predicted and gold culture-specific item spans under one-to-one alignment). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 60.47 | #455 |
| 2 | Qwen 2.5 Max | 59.75 | #962 |
| 3 | GPT-4o | 57.74 | #741 |
| 4 | OLMo 2 32B | 57.71 | #5193 |
| 5 | Gemini 2.0 Flash | 57.15 | #702 |
| 6 | DeepSeek R1 Distill Qwen 32B | 55.14 | #1780 |
| 7 | Llama 3.3 70B Instruct | 48.76 | #1353 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=xcr-bench-csi-identification · How It Works · Data refreshed daily, snapshot 2026-11-12.