We pitted LLMs against themselves in different languages across games🎮 Usually when a LLM is worse at a task in some language, we go "it's a language issue". But isn't that circular? If the task is made of language, how do you tell if the model lacks skills, or just struggles to read the question? 🧵

At last, a way to unlock mulitlingual knowledge sharing in LLMs! 🌍 ​By pretraining an English/Arabic model and swapping Arabic for a word-wise, 1-to-1 translation mapping to English, we saw a massive boost in cross-lingual knowledge transfer 🚀

LLMs can't generalize across 2 copies of the same language!🫠 Usually we blame multilinguality issues on syntax, tokenizer fragmentation or data disproportion, so what if we eliminate these factors? In our ongoing work by Adam Jaber Bobby Cheng and I we ran a pretraining experiment doing just that