Why isn’t LLM reasoning done in vector space instead of natural language?[D]

About this article

Why don’t LLMs use explicit vector-based reasoning instead of language-based chain-of-thought? What would happen if they did? Most LLM reasoning we see is expressed through language: step-by-step text, explanations, chain-of-thought style outputs, etc. But internally, models already operate on high-dimensional vectors. So my question is: Why don’t we have models that reason more explicitly in latent/vector space instead of producing intermediate reasoning in natural language? Would vector-bas...