Google

An alternative approach to optimizing soft prompts involves minimizing the Kullback-Leibler (KL) divergence between the output probability distribution from the full context, $$\text{Pr}(\cdot|\mathbf{c}, \mathbf{z})$$, and the distribution from the soft prompt, $$\text{Pr}(\cdot|\sigma, \mathbf{z})$$. The goal is to find the soft prompt $$\hat{\sigma}$$ that makes these two distributions as similar as possible. The optimization is expressed by the formula: $$\hat{\sigma} = \underset{\sigma}{\arg\min}\, \text{KL}(\text{Pr}(\cdot|\mathbf{c}, \mathbf{z}) \|\| \text{Pr}(\cdot|\sigma, \mathbf{z}))$$

Formula for Soft Prompt Optimization by Minimizing KL Divergence

Based on the optimization objective of minimizing the KL divergence between the distribution from the full context and the distribution from the soft prompt, which soft prompt (σ_A or σ_B) is superior? Justify your answer by explaining how the distributions relate to the optimization goal, without calculating the exact KL divergence value.

Evaluating Soft Prompt Performance

Consider the formula for finding an optimal soft prompt, $\hat{\sigma}$, by minimizing the difference between two probability distributions: $$\hat{\sigma} = \underset{\sigma}{\arg\min}\, \text{KL}(\text{Pr}(\cdot|\mathbf{c}, \mathbf{z}) \|\| \text{Pr}(\cdot|\sigma, \mathbf{z}))$$ In this formula, $\text{Pr}(\cdot|\mathbf{c}, \mathbf{z})$ is the probability distribution over possible outputs given a full context $\mathbf{c}$ and an input $\mathbf{z}$, while $\text{Pr}(\cdot|\sigma, \mathbf{z})$ is the distribution given a soft prompt $\sigma$ and the same input $\mathbf{z}$.

Explain why $\text{Pr}(\cdot|\mathbf{c}, \mathbf{z})$ is treated as the first argument (the 'true' distribution) and $\text{Pr}(\cdot|\sigma, \mathbf{z})$ as the second argument within the KL divergence function, and not the other way around. What would be the conceptual implication of swapping their positions?

Analyzing the Asymmetry in Soft Prompt Optimization

A researcher is training a soft prompt, denoted as $\sigma$, to mimic the behavior of a full context, $c$, for a given input, $z$. They use the Kullback-Leibler (KL) divergence between the model's output probability distributions as their objective function: $$\text{KL}(\text{Pr}(\cdot|c, z) \|\| \text{Pr}(\cdot|\sigma, z))$$ After extensive training, the researcher observes that the KL divergence has reached a value of 0. What is the most accurate conclusion to draw from this result?

Learn Before

Related