Google

Pre-trained Large Language Models are fundamentally optimized for next-token prediction, meaning their primary training objective is to predict the next word in a sequence. This training method does not inherently equip them to follow explicit instructions. Consequently, when presented with a task-oriented prompt, such as a request for summarization, a pre-trained LLM is likely to simply continue the input text rather than executing the given command.

Limitation of Pre-trained LLMs: Next-Token Prediction vs. Instruction Following

Instruction alignment, also known as instruction fine-tuning, is the process of adapting a Large Language Model to accurately follow user instructions and intentions. This tuning addresses the core limitation of pre-trained models, which are optimized for next-token prediction and thus tend to continue input text rather than executing commands. Key challenges and areas of focus within instruction alignment include the methods for fine-tuning, the generation and collection of high-quality instruction data, and ensuring the model can generalize to new, unseen instructions.

Instruction Alignment

Instead of providing a summary, the model generates the following text: '...which consists of the Sun and the objects that orbit it, including eight planets and their moons.' Based on the model's described training, explain the fundamental reason for the discrepancy between the developer's instruction and the model's actual output.

Diagnosing Language Model Output

A large language model has been developed with a single training objective: to become highly proficient at predicting the most probable next word in a sequence of text. If a user provides this model with the input, 'Write a short story about a robot who discovers music.', what kind of output is the model most likely to produce? Explain your reasoning based on its training objective.

Predicting Pre-trained Model Behavior

A user interacts with a large language model that has only undergone its initial training phase on a vast corpus of text, without any subsequent fine-tuning to follow commands. The user provides the input: 'Translate the following sentence into French:'. Which of the following outputs is most characteristic of this specific type of model's behavior?

Learn Before

Related