TLDR: A new AI framework called Dual-Stage Refinement (DSR) improves screenplay generation by decoupling creative storytelling from strict formatting. It first converts an outline into novel-style prose, then refines that prose into a professional screenplay. A hybrid data synthesis method trains the model, leading to screenplays that achieve a 75% win rate against top LLMs and 82.7% of human quality in expert evaluations.
The research paper “Beyond Direct Generation: A Decomposed Approach to Well-Crafted Screenwriting with LLMs” by Hang Lei, Shengyi Zong, Zhaoyan Li, Ziren Zhou, and Hao Liu from Alibaba Group and Peking University, explores a new method for Large Language Models (LLMs) to generate high-quality screenplays. The authors argue that directly generating screenplays from start to finish often results in outputs that look good on the surface but lack the deep structure and compelling storytelling needed for professional use. This is because a single model struggles to simultaneously handle both creative narrative construction and the strict formatting rules of screenwriting.
To address this challenge, the researchers introduce a framework called Dual-Stage Refinement (DSR). Instead of a single, direct generation process, DSR breaks down screenplay creation into two distinct stages. The first stage focuses purely on creative narrative development. Here, a brief outline is transformed into rich, novel-style prose. This allows the model to concentrate solely on the storytelling aspect, developing plot, character interactions, and emotional arcs without worrying about formatting.
The second stage then takes this narrative prose and refines it into a professionally formatted screenplay. This stage is dedicated exclusively to stylistic conversion, ensuring that the text adheres to the specific conventions of screenwriting, such as scene headings, concise action lines, and authentic dialogue. By separating these two complex tasks, the DSR framework enables the LLM to specialize in one distinct capability at each stage, leading to better overall performance.
A significant hurdle in implementing DSR was the lack of suitable training data, specifically paired outline-to-novel texts required for the first stage. The team tackled this with an innovative hybrid data synthesis strategy. This process begins with “reverse synthesis,” where existing screenplays are deconstructed into structured inputs like outlines and character profiles. Following this, “forward synthesis” uses these deconstructed inputs to generate high-quality, narratively rich novel-style texts, which then serve as training targets for the first stage of the DSR model. This tailored data creation process is crucial for the framework’s effectiveness.
Also Read:
- AI Learns Your Story Preferences for Truly Personalized Narratives
- Simulating Human Surveys: A New Approach for LLMs to Capture Preference Shifts
Extensive experiments were conducted, including blind evaluations by professional screenwriters. The results were impressive: screenplays generated by the DSR framework achieved a 75% win rate against strong baseline models like Gemini-2.5-Pro and Claude-Sonnet-4. Furthermore, the quality of the generated scripts reached 82.7% of human-level performance. This demonstrates that a decomposed generation architecture, supported by a specialized hybrid data synthesis strategy, is a highly effective approach for training LLMs in complex creative fields. The DSR framework significantly improves LLMs’ ability to manage narrative structure, character development, and dramatic expression, bringing AI-assisted screenwriting closer to professional standards. You can read the full research paper for more details here.
