Steering Large Language Models to Evaluate and Amplify Creativity

TL;DR

This paper presents a method to evaluate and enhance the creativity of large language models by analyzing their internal states when generating creative versus boring responses, aligning better with human judgment.

Contribution

It introduces a mechanistic approach to measure and improve LLM creativity by leveraging internal state differences during response generation.

Findings

Internal state differences correlate with human judgments of creativity.

The method improves the creativity of generated text at inference time.

The approach provides a robust measure of creativity for LLMs.

Abstract

Although capable of generating creative text, Large Language Models (LLMs) are poor judges of what constitutes "creativity". In this work, we show that we can leverage this knowledge of how to write creatively in order to better judge what is creative. We take a mechanistic approach that extracts differences in the internal states of an LLM when prompted to respond "boringly" or "creatively" to provide a robust measure of creativity that corresponds strongly with human judgment. We also show these internal state differences can be applied to enhance the creativity of generated text at inference time.