StoryLens: AI Benchmark for Preference-Aligned Story Rewriting
A team of researchers has unveiled StoryLensBench, an extensive benchmark designed for preference-aligned story rewriting, along with StoryLensEval, a model that assesses reader satisfaction, and StoryLensWriter, a two-phase rewriting system. In a preliminary human study, results indicated that merely adapting style results in a modest 2.3% boost in reader satisfaction, whereas rewriting with enhanced context leads to a significant 24.5% increase. The benchmark includes structured storybooks, diverse reader preference profiles, and ranked context-aware rewritten narratives. This research is documented on arXiv (2605.28073) and emphasizes the importance of enriching narratives beyond just superficial stylistic changes.
Key facts
- StoryLensBench is a large-scale benchmark for preference-aligned story rewriting
- StoryLensEval is a reward model for estimating reader satisfaction
- StoryLensWriter is a two-stage rewriting model combining supervised learning
- Style adaptation alone provides only 2.3% improvement in reader satisfaction
- Context-enhanced rewriting improves user preference alignment by 24.5%
- The benchmark includes structured story books, preference profiles, and rewritten stories
- The research is published on arXiv with ID 2605.28073
- The work argues for context-aware narrative enrichment beyond style transfer
Entities
Institutions
- arXiv