The probability distribution over possible next tokens that a language model produces during decoding.