Latent-thought paper formalizes four axioms for LLM reasoning representations

A new arXiv paper proposes four axioms for formalizing latent thoughts in large language models. The work attempts to make internal reasoning representations more explicit and easier to discuss.

The topic matters because many debates about reasoning in LLMs rely on vague language about thoughts, chains and hidden states. A clearer formal framework could help researchers compare methods and claims.

The paper is part of a broader effort to understand whether model reasoning can be represented, inspected or shaped without relying only on visible chain-of-thought text.