Learn Before
Choosing an Optimization Strategy for Soft Prompts
A research lab is working on two distinct projects that use soft prompts to compress lengthy contextual information. For each project, recommend one of the two primary optimization strategies (maximizing the log-probability of the desired output OR minimizing the divergence between the full-context and soft-prompt output distributions). Justify your choice for each project.
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Evaluation in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
Formula for Soft Prompt Optimization via Log-Likelihood Maximization
Formula for Soft Prompt Optimization by Minimizing KL Divergence
A team is creating a soft prompt to summarize a complex user manual for a question-answering model. Their main objective is not just to get the single correct answer, but to ensure the model's uncertainty and its ranking of other plausible-but-incorrect answers are the same with the soft prompt as they were with the full manual. Which of the following optimization strategies best aligns with this specific objective?
Choosing an Optimization Strategy for Soft Prompts
A researcher is optimizing a soft prompt. With the original, long context, the model predicts the correct answer with 60% probability and a plausible alternative with 30% probability. The researcher's goal is to create a soft prompt that causes the model to predict the correct answer with over 95% probability, even if this significantly changes the probability of the alternative answer. Which optimization approach is better suited for this specific goal?