Retell, Reward, Repeat: Reinforcement Learning for Narrative Theory-Informed Story Retelling

The paper introduces "Retell, Reward, Repeat" (RRR), a reinforcement learning pipeline that combines Structuralist Narratology with scalar narrativity to enhance storytelling in LLMs. RRR utilizes an extended TimeTravel dataset with human-annotated narrative stages and employs d-RLAIF to derive training signals from the narrativity of textual features, achieving superior performance in logic, rationality, and completeness compared to few-shot and SFT baselines. This approach emphasizes the importance of integrating linguistic theories into NLP, offering a cost-effective post-training method for narrative generation in AI systems.