Latent Reasoning
Letting a model reason using its own internal hidden representations instead of writing every reasoning step out as words.
Definition
Latent reasoning has a model work through a problem using its internal, continuous hidden states instead of writing out each reasoning step as natural-language tokens. A notable example is Coconut (Chain of Continuous Thought), which feeds a model's own last hidden state back in as the next input, skipping the step where it has to be decoded into a word. Because a single hidden state can blend several possible next steps at once, this lets the model keep multiple lines of reasoning in play simultaneously rather than committing to one word-by-word path, and it can use fewer tokens than writing everything out. The tradeoff is that this internal reasoning isn't directly readable by a person the way a written chain of thought is, which has implications for how easily outside observers can check why the model reached its answer.