StoryLens: Preference-Aligned Story Rewriting via Context-Aware Narrative Enrichment

Researchers introduce StoryLens, a framework for preference-aligned story rewriting that goes beyond style transfer to incorporate context-aware narrative enrichment. Human studies show context-enhanced rewriting improves reader satisfaction by 24.5% compared to style-only approaches, supported by a new benchmark, reward model, and two-stage rewriting system combining supervised learning with reinforcement learning.

StoryLens addresses a fundamental gap in narrative AI: the distinction between surface-level stylistic changes and meaningful, context-aware story adaptation. The research demonstrates that traditional style transfer approaches—which focus solely on tone, vocabulary, and grammatical patterns—deliver minimal improvements in reader satisfaction at just 2.3%. This finding challenges the assumption that story personalization is primarily a style problem, revealing instead that preserving and enriching narrative context substantially matters to readers.

The work builds on decades of NLP research in style transfer and personalization, but applies these techniques specifically to long-form narrative. Prior work in story generation and style adaptation operated at smaller scales or focused on simplified metrics like readability scores. StoryLens introduces STORYLENSBENCH as a structured dataset with preference profiles and ranked outputs, establishing benchmarking standards that the community currently lacks. This mirrors the maturation seen in other NLP domains where well-designed benchmarks accelerated progress.

For AI developers and research institutions, this work has direct implications. The combination of STORYLENSEVAL (a satisfaction reward model) and STORYLENSWRITER (the two-stage generation system) provides a reproducible methodology for preference-aligned generation tasks beyond storytelling. The use of GRPO-based reinforcement learning alongside supervised fine-tuning represents a practical approach to aligning language models with nuanced human preferences, relevant across content creation, personalization systems, and creative AI applications.

Looking forward, the significance lies in whether these techniques generalize to other narrative domains—screenwriting, game narratives, interactive fiction—and whether preference models trained on this data can scale efficiently. The research also opens questions about computational cost and the practical deployment of reward models in production systems.

• →Context-aware narrative enrichment increases reader satisfaction 10x more than style adaptation alone (24.5% vs 2.3%)
• →StoryLens introduces the first large-scale benchmark specifically designed for preference-aligned story rewriting with structured preference profiles
• →Two-stage approach combining supervised fine-tuning with GRPO reinforcement learning outperforms standard generation and personalization baselines
• →Reward models for estimating reader satisfaction over narratives represent a scalable approach to preference alignment in creative AI
• →The distinction between style transfer and context-aware narrative enrichment suggests previous story rewriting approaches were fundamentally misaligned with user needs