Which parameter in a large language model (LLM) determines the number of potential next tokens evaluated during each step of text generation?