LLMs require sizable text corpora for training, which is unevenly distributed across languages, thus affecting model performance. Li et al. propose the Language Ranker to measure LLMs’ performance across a variety of languages, benchmarking them against English. Their study reveals that LLM performance ranking is consistent across languages and that model size does not affect the language performance hierarchy.

• Proposed Language Ranker benchmarks LLMs’ multilingual performance.
• LLMs exhibit a stable performance order across different languages.
• Correlation between language performance and pre-training corpus proportion.

The insights from this study shed light on the discrepancies in LLM performances and encourage efforts to boost AI’s linguistic inclusivity. Discover More