SWAG: Storytelling With Action Guidance

SWAG improves long-form story generation using two-model feedback loop, outperforming previous techniques.

Summary:

• SWAG is a novel approach to storytelling with large language models (LLMs) that reduces story writing to a search problem through a two-model feedback loop.
• The SWAG pipeline using only open-source models surpasses GPT-3.5-Turbo in terms of performance.
• The SWAG feedback loop can be run as many times as needed until the desired story length is reached.

Major Findings:

1. SWAG substantially outperforms previous end-to-end story generation techniques when evaluated by GPT-4 and through human evaluation.
2. The SWAG pipeline using only open-source models surpasses GPT-3.5-Turbo.
3. The SWAG feedback loop can be run as many times as needed until the desired story length is reached.

Analysis and Critique:

• The article provides a comprehensive overview of the SWAG approach to storytelling with LLMs, highlighting its effectiveness in generating engaging and captivating stories.
• The use of human and machine evaluations demonstrates the superiority of SWAG over end-to-end story generation techniques.
• The limitations of the study include the use of DPO for AD LLM alignment due to compute restraints, as well as the constraints in evaluating a larger set of stories for both machine and human evaluations.

Appendix

Model	gpt-3.5-turbo-1106
Date Generated	2024-02-26
Abstract	https://arxiv.org/abs/2402.03483v1
HTML	https://browse.arxiv.org/html/2402.03483v1
Truncated	False
Word Count	7252