Language-Agnostic Representations Show a Shared Semantic Workspace

The standard story is that LLMs work in words. They predict the next token, so surely their internal...