💬 NLP

Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion

This paper introduces an attribute-guided genre expansion framework that scales creative writing data beyond stories by combining human prompts with curated genre attributes to generate a 50K-example Multi-Genre Collection, which significantly enhances LLMs' ability to master diverse creative formats and structural conventions.

Original authors: Hwan Chang, Yongil Kim, Heuiyeen Yeen, Yireun Kim, Jinsik Lee, Hwanhee Lee

Published 2026-08-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hwan Chang, Yongil Kim, Heuiyeen Yeen, Yireun Kim, Jinsik Lee, Hwanhee Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot how to be a writer. For a long time, the best way to do this was to feed the robot millions of stories—adventures, mysteries, and fairy tales. This worked great for teaching the robot how to tell a tale with a beginning, middle, and end. But in the real world, writing isn't just about stories. Sometimes you need to write a rap song with a specific rhyme scheme, a script for a video game with strict rules, or a news article with a very specific format. If you only feed the robot stories, it might become a great storyteller but a terrible rapper or game designer because it doesn't know the "rules of the road" for those other formats. This paper tackles that exact problem: how to teach AI to write in many different styles, not just one.

The researchers behind this study, Hwan Chang and his team, realized that simply giving the AI more stories wouldn't fix the problem. It's like trying to learn how to play basketball just by watching a thousand movies about soccer; you'll understand the concept of a game, but you won't know how to dribble or shoot. They needed a way to teach the AI the specific "structural DNA" of different creative formats. To do this, they built a new system called Attribute-Guided Genre Expansion. Think of it as a creative factory. Instead of just handing the AI a blank page and saying "write something," they give it two things: a "seed" (a cool idea or theme, like "a cyberpunk city") and a "rulebook" (a set of specific constraints, like "use AABB rhymes" or "include timestamps"). By mixing and matching these seeds and rulebooks, they created a massive library of 50,000 examples covering 13 different creative genres, from movie scripts to technical manuals.

The paper suggests that this method works incredibly well. When they trained AI models on this new, diverse library, the models didn't just get better at writing stories; they became much better at following the strict rules of other formats too. In fact, these models outperformed other AI models that had been trained on existing, specialized writing datasets. The researchers found a clear pattern: the more different types of writing the AI learned, the more creative and unique its outputs became. It's as if the AI learned that a "story" and a "rap song" are different languages, and by learning both, it became a more fluent speaker overall.

However, the paper also points out what doesn't work. They argue that simply scaling up the amount of story data is not the answer. If you just feed an AI more stories, it might get better at describing a dragon, but it will still fail to format a game design document correctly because it hasn't learned the specific structural constraints of that genre. The study explicitly rules out the idea that "more of the same" is the solution. Instead, they suggest that the key to unlocking true creative writing ability in AI is variety and structure.

The team built their dataset by taking human-written story prompts (like those from a popular online writing community) and using a smart process to transform them into different genres. They used a "thematic seed" (the core idea) and injected "genre attributes" (the specific rules) to generate new examples. For instance, they took a prompt about "urban loneliness" and turned it into a rap verse with a specific flow, a movie script with dialogue rules, and a podcast script with a specific tone. They then used powerful AI models to write the responses and carefully filtered out the bad ones to ensure high quality.

The results were measured using several tests. The AI models trained on this new "Multi-Genre Collection" consistently scored higher than models trained on just stories or other existing writing datasets. They performed better on general writing challenges and were more creative. The researchers also tested how "novel" the AI's writing was. They found that as they increased the number of genres the AI was trained on (from just stories to all 13 genres), the AI's ability to produce fresh, non-repetitive ideas improved steadily. This suggests that exposure to diverse formats directly boosts the model's creative flexibility.

In short, this paper proposes that to make AI a true creative partner, we need to stop treating it like a storybook machine and start treating it like a multi-talented artist. By teaching it the specific rules of 13 different creative worlds, we don't just get better stories; we get better raps, better scripts, and better game designs. The authors conclude that this approach of mixing thematic ideas with strict genre rules is a powerful way to scale up creative writing capabilities, offering a path forward for AI that is as versatile as human writers.

Technical Summary: Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion

Problem Statement

Current high-quality creative writing datasets for Large Language Models (LLMs) are heavily dominated by story-centric data. While effective for narrative generation, this limitation hinders models from adhering to the distinct structural, stylistic, and functional conventions of diverse creative formats (e.g., rap lyrics, game design documents, movie scripts). Existing synthetic instruction generation methods often fail in this domain because creative writing lacks compact validity signals (like code correctness or math answers) to guide automated scaling. Consequently, simply scaling story data does not address the "structural gaps" required for genres that demand specific formatting, rhyme schemes, or mechanical rules.

Methodology

The authors propose an Attribute-Guided Genre Expansion framework to construct the Multi-Genre Collection, a 50,000-example corpus spanning 13 creative genres. The methodology decouples thematic breadth from genre-form control through a three-stage pipeline:

Thematic Seed Sampling:
- The process begins with a curated pool of human-authored story prompts from r/WritingPrompts.
- After safety and relevance filtering (using GPT-5-mini), 5 diverse query-response pairs are sampled per iteration to serve as "thematic seeds." These provide topical and stylistic variety without imposing story-specific constraints.
Genre Attribute Sampling:
- For each of the 13 target genres (including Rap, Movie Script, Game Design, Lyrics, etc.), a set of 5–15 structural attributes is manually curated from encyclopedic references and writing manuals (e.g., "Rhyme Scheme" and "Flow" for rap; "Narrative Arc" for scripts).
- To simulate natural variation in instruction specificity, a subset of $k$ attributes is randomly sampled from the genre's attribute set ( $k$ ranges from 0 to the total number of attributes).
- This step ensures that generated queries vary in constraint levels, from open-ended to highly specific.
Attribute-Guided Pair Synthesis:
- A meta-prompt template combines the target genre, the sampled thematic seeds, and the sampled attributes.
- Query Generation: GPT-5-mini generates diverse writing queries based on the meta-prompt, instructed to vary in topic, tone, and structure.
- Response Generation: A strong model (Qwen3-235B-A22B-Thinking) generates the corresponding creative content.
- Quality Filtering: An independent LLM-as-a-judge (Qwen3-30B-A3B-Instruct) evaluates response quality. Pairs scoring below two standard deviations from the mean are discarded.

The resulting corpus is balanced across 13 primary genres, with an additional "Etc." category for long-tail formats, totaling 50K instances.

Key Contributions

Multi-Genre Collection: A new 50K-instance dataset covering 13 distinct creative writing genres, moving beyond the story-centric paradigm.
Attribute-Guided Framework: A novel data synthesis method that separates thematic diversity (derived from human seeds) from structural fidelity (enforced by curated genre attributes). This approach addresses the lack of validity signals in creative writing by explicitly encoding genre conventions.
Genre-Count Ablation: An analysis demonstrating that increasing the number of training genres directly correlates with improved output novelty and diversity.

Experimental Results

The authors fine-tuned three base models (Llama-3.1-8B, EXAONE-3.5-7.8B, and Qwen3-8B) on the Multi-Genre Collection and evaluated them against base models, writing-specialized baselines (LongWriter-glm4-9B), and models trained on existing writing corpora (DeepWriting, LongWriter).

Performance Gains: Models fine-tuned on the Multi-Genre Collection consistently outperformed base models and the writing-specialized LongWriter baseline across all benchmarks (Arena Hard, WritingBench, and an in-distribution Multi-Genre test set).
Comparison with Existing Data: When trained on equal-sized subsets (2K examples), the Multi-Genre Collection yielded superior results compared to DeepWriting and LongWriter datasets, particularly in out-of-distribution generalization (Arena Hard).
Impact of Genre Diversity: Experiments varying the number of training genres (from 0 to 13) showed a monotonic increase in the "Novelty" metric, confirming that exposure to diverse formats enhances the model's ability to generate non-redundant, varied outputs.
Robustness: Improvements were validated using independent LLM judges (DeepSeek, Gemini) and human evaluation, ruling out evaluator bias or self-preference artifacts.
Qualitative Improvements: Case studies showed that fine-tuned models produced outputs with better format compliance (e.g., Suno-compatible timestamps), specific world-building (named entities/settings), and narrative coherence compared to base models.

Significance

The paper claims that the Multi-Genre Collection and the attribute-guided expansion framework effectively bridge the gap between story-centric training data and the broader requirements of real-world creative writing. The authors assert that genre expansion is a key driver of robust creative writing capability, suggesting that diverse, genre-faithful data is more effective for general creative ability than narrow specialization on a single format. The work provides a scalable, controlled methodology for generating high-quality synthetic data for domains where traditional validity signals are absent.