Prompt Waywardness: The Curious Case of Discretized Interpretation of Continuous Prompts

TL;DR

This paper explores the challenge of interpreting continuous prompts in language models by examining their discretized counterparts, revealing surprising 'wayward' behaviors and implications for faithful interpretation and generalization.

Contribution

It uncovers the phenomenon of waywardness in discretized prompts, providing empirical analysis and insights into the behavior of prompts across model sizes and tasks.

Findings

Larger models exhibit higher waywardness, mapping prompts more closely to arbitrary texts.

Continuous prompts can solve tasks while being projected to contradictory or unrelated texts.

Faithful interpretation of continuous prompts is fundamentally challenging due to observed behaviors.

Abstract

Fine-tuning continuous prompts for target tasks has recently emerged as a compact alternative to full model fine-tuning. Motivated by these promising results, we investigate the feasibility of extracting a discrete (textual) interpretation of continuous prompts that is faithful to the problem they solve. In practice, we observe a "wayward" behavior between the task solved by continuous prompts and their nearest neighbor discrete projections: We can find continuous prompts that solve a task while being projected to an arbitrary text (e.g., definition of a different or even a contradictory task), while being within a very small (2%) margin of the best continuous prompt of the same size for the task. We provide intuitions behind this odd and surprising behavior, as well as extensive empirical analyses quantifying the effect of various parameters. For instance, for larger model sizes we observe higher waywardness, i.e, we can find prompts that more closely map to any arbitrary text with a smaller drop in accuracy. These findings have important implications relating to the difficulty of faithfully interpreting continuous prompts and their generalization across models and tasks, providing guidance for future progress in prompting language models.