Robust Preference Learning for Storytelling via Contrastive Reinforcement Learning

TL;DR

This paper introduces a contrastive reinforcement learning approach to improve controlled story generation, creating a robust preference model that aligns stories with human critiques and enhances generation quality.

Contribution

It develops CARP, a contrastive bi-encoder for preference modeling, and demonstrates its effectiveness in fine-tuning story generation with improved robustness and human preference alignment.

Findings

Full pipeline outperforms larger LLMs and logit-based methods in human preference tests

Contrastive reward modeling enhances story generation robustness

Human study confirms improved quality of generated stories

Abstract

Controlled automated story generation seeks to generate natural language stories satisfying constraints from natural language critiques or preferences. Existing methods to control for story preference utilize prompt engineering which is labor intensive and often inconsistent. They may also use logit-manipulation methods which require annotated datasets to exist for the desired attributes. To address these issues, we first train a contrastive bi-encoder model to align stories with corresponding human critiques, named CARP, building a general purpose preference model. This is subsequently used as a reward function to fine-tune a generative language model via reinforcement learning. However, simply fine-tuning a generative language model with a contrastive reward model does not always reliably result in a story generation system capable of generating stories that meet user preferences. To increase story generation robustness we further fine-tune the contrastive reward model using a prompt-learning technique. A human participant study is then conducted comparing generations from our full system, ablations, and two baselines. We show that the full fine-tuning pipeline results in a story generator preferred over a LLM 20x as large as well as logit-based methods. This motivates the use of contrastive learning for general purpose human preference modeling.