Google

A key distinction between different scaling strategies for Large Language Models lies in their mechanism for improvement. Pre-training and fine-tuning scaling rely on updating the model's parameters through further training, whereas inference-time scaling enhances model performance during application without altering its trained parameters.

Comparison of Scaling Approaches: Parameter Updates vs. Inference-Time Methods

A tech startup wants to improve its general-purpose language model for the specific task of summarizing customer support tickets. Two teams propose different methods. Based on the descriptions provided in the case study, which team's proposal involves modifying the model's underlying parameters, and which one enhances performance at the moment of use without such modifications? Justify your answer by explaining the core difference in their approaches.

Evaluating Model Improvement Strategies

Consider two common methods for improving a Large Language Model's performance: (1) Domain-adaptive fine-tuning and (2) Chain-of-Thought prompting. For each method, identify whether it relies on parameter updates or is an inference-time method. Justify your classification for both.

Classifying LLM Scaling Strategies

A development team is using a large, pre-trained language model that is computationally expensive to modify. They need to enhance its performance for a specific, temporary project. A key requirement is that any performance enhancement must be easily removable, restoring the model to its original state without needing to store a separate version. Which scaling approach is most suitable for this scenario?

Learn Before

Related