Which best captures the key idea of "How LLMs generate text"?

ABuild review checklists for a regulation area BThe model assigns probabilities to possible next tokens and samples one, repeating until the... CStore facts outside the prompt and re-inject only what's relevant DCache common results and cap lengths to control spend