Google

A key difference exists in the difficulty of optimizing prompt instructions versus demonstrations. Generating high-quality demonstrations with LLMs is relatively straightforward, shifting the optimization challenge to sampling the best examples from a candidate pool. Conversely, learning instructions is harder because LLMs struggle to predict instruction quality, and the required evaluation process is computationally expensive.

Comparison of Optimizing Prompt Instructions vs. Demonstrations

When designing prompts for a large language model, why is the process of optimizing the *instructions* generally considered a more significant computational challenge than optimizing the *demonstrations* (examples)?

A machine learning team is working to improve the performance of a large language model on a new task. They have a very limited computational budget. They are considering two different approaches to enhance their prompt. Based on the typical challenges of prompt engineering, which of the following two strategies is more likely to be a cost-effective and successful use of their limited resources? Justify your reasoning.

Prompt Optimization Strategy Selection

When automatically optimizing a prompt, the main challenge associated with the demonstrations (examples) is the difficulty of generating high-quality candidates, whereas the main challenge for the instructions is the difficulty of sampling the best ones from a large generated pool.

Learn Before

Related