Toward Human Readable Prompt Tuning: Kubrick's The Shining is a good movie, and a good prompt too?

TL;DR

This paper explores what makes natural language prompts effective for large language models, proposing a human-readable prompt tuning method that leverages fluency and topical relevance, and demonstrating improved performance with unlabeled data.

Contribution

It introduces FLUENT PROMPT, a prompt tuning approach using Langevin dynamics with fluency constraints, and shows how effective prompts relate to task topics and label word probabilities.

Findings

Effective prompts are topically related to tasks.

Prompts calibrate prior probabilities of label words.

Unsupervised prompt generation outperforms baselines by 7% accuracy.

Abstract

Large language models can perform new tasks in a zero-shot fashion, given natural language prompts that specify the desired behavior. Such prompts are typically hand engineered, but can also be learned with gradient-based methods from labeled data. However, it is underexplored what factors make the prompts effective, especially when the prompts are natural language. In this paper, we investigate common attributes shared by effective prompts. We first propose a human readable prompt tuning method (F LUENT P ROMPT) based on Langevin dynamics that incorporates a fluency constraint to find a diverse distribution of effective and fluent prompts. Our analysis reveals that effective prompts are topically related to the task domain and calibrate the prior probability of label words. Based on these findings, we also propose a method for generating prompts using only unlabeled data, outperforming strong baselines by an average of 7.0% accuracy across three tasks.