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

A viral Reddit post by user ZeusZCC has sparked a deep debate among AI researchers and engineers: Why do LLMs reason in natural language via chain-of-thought (CoT) text, when internally they operate on high-dimensional vectors? The post questions whether explicit vector-based reasoning would be faster and more compressed, allowing models to 'think' in latent space and only translate the final output into language. This could potentially unlock more intuitive, human-like reasoning without the verbosity of step-by-step text.

However, critics argue that vector-based reasoning would make AI outputs opaque and difficult to verify, especially for domains requiring strict logic like mathematics, programming, and legal reasoning. Chain-of-thought reasoning, while slower, provides a transparent audit trail that humans can inspect for errors. The debate highlights a fundamental trade-off: vector reasoning might be more efficient but at the cost of interpretability, which is crucial for trust and debugging in high-stakes applications. Researchers are now exploring hybrid approaches that combine both methods.