New Framework Enhances Controllable Narrative Generation for Creative Writing

Large language models often struggle with creative writing, either offering superficial edits or making uncontrolled, destructive plot changes. This creates a fundamental tension between maintaining the original story's integrity and enhancing its descriptive richness. Researchers have developed Loom, a new framework designed to overcome this challenge. Loom operates on a three-layer pipeline that separates the generation of perceptual details from their syntactic insertion. It employs an intent-centered semiotic chain-of-thought, allowing for precise control over the narrative's intended meaning and the density of its descriptive rendering. This architecture ensures that creative enhancements are made without altering the core event structure of the original text. Evaluations, including both LLM-based metrics and human assessments, show that Loom significantly improves factual integrity and descriptive intensity compared to existing methods, achieving the highest overall quality scores. This indicates a successful resolution of the long-standing trade-off in AI-assisted creative writing.