Evaluating and Improving Multilinguality in Large Language Models

I'm interested in making LLMs better at multilinguality. There are several dimensions to this problem:

• 01 Evaluating or improving — We could focus on evaluating existing capabilities, or on improving them.
• 02 Comprehension or generation — Understanding multilingual inputs and generating multilingual outputs are different goals, often needing different kinds of approaches.
• 03 Resourcedness of language — Some languages are practically unseen to models with little or no data, others are low-resource with very little data, and then there are mid-resource languages with decent existing LLM support. Many low-resource languages also have dialectal relationships with related high-resource languages, which can be exploited in several ways.
• 04 Techniques — Finally, there are a number of ways to tackle these problems.

Here's a schema of my work categorised along the above dimensions! These are first author works, except for [4] and [9], in which I mentored students as the last author. [8] was a large team effort at Meta with 9 core contributors including me, for which I led the component on post-training dataset design.