“Despite research into latent space reasoning, successful reasoning models from OpenAI, Anthropic, and Google primarily perform chain-of-thought in the standard token space using human language.”