A researcher pre-trains a new language model on a vast and diverse dataset of text from the internet. Without any subsequent specialized training, the researcher tests the model with the prompt: 'Summarize the following paragraph in one sentence: [paragraph text]'. The model successfully produces a coherent, one-sentence summary. Which of the following statements provides the most accurate explanation for this capability?

Zero-Shot Generalization from Pre-trained Instruction Knowledge

A large language model's capacity to understand and execute a wide range of tasks based on textual prompts is primarily instilled through a specialized training stage that is separate from and follows its initial, general-purpose language learning phase.

Impact of Pre-training Data on Instruction Following