High-quality creative writing data for large language models (LLMs) remains dominated by story-centric data, limiting models' ability to follow the structural and functional conventions of diverse creative formats. We propose an attribute-guided genre expansion framework for scaling creative writing data beyond story generation. By separating thematic breadth from genre-form control, our framework leverages human-authored story prompts as diverse creative seeds, while utilizing manually curated genre attributes to enforce distinct structural, stylistic, and formatting conventions. We combine these to prompt strong LLMs for genre-faithful query-response pairs, which are then quality-filtered. Applying this framework, we construct the Multi-Genre Collection, a 50K-example corpus spanning 13 creative genres, including story, rap, lyrics, scripts, game design, character design, and other creative formats. Experiments across out-of-distribution writing benchmarks and held-out genre diagnostics demonstrate that models fine-tuned on our data consistently surpass not only base models and writing-specialized baselines, but also models trained on existing writing corpora. Genre-count ablations further indicate that controlled genre expansion, rather than story-centric scaling alone, is a key driver of robust creative writing capability.
Figures & tables
Figure 1. Overview of our attribute-guided multi-genre expansion framework. A thematic seed is transformed into diverse outputs by injecting curated genre attributes. Overview diagram showing how thematic seeds are transformed into genre-faithful outputs through curated genre attributes.
Figure 2. Attribute-guided genre expansion pipeline. We sample few-shot examples from existing story prompts to maintain thematic diversity, randomly sample subset genre-specific attributes to inject structural constraints, and generate multi-genre queries by combining task instructions, few-shot demonstrations, target genre, and sampled attributes. Pipeline diagram illustrating few-shot sampling, genre-attribute injection, and query generation for attribute-guided multi-genre data construction.
Figure 4. Genre distribution in the Multi-Genre Collection. Distribution chart showing that the Multi-Genre Collection contains balanced coverage across 13 creative writing genres.
Figure 5. t-SNE visualization of the Multi-Genre Collection. t-SNE scatter plot showing semantic clusters for the 13 genres in the Multi-Genre Collection.
OOD (Generalization)
ID (Coverage)
Models
Arena Hard
WritingBench
Multi-Genre
LongWriter
2.3
47.2
49.5
Llama
2.9
43.7
44.4
Llama + SFT
33.0 (+30.1)
60.6 (+16.9)
68.4 (+24.0)
EXAONE
10.4
52.7
64.9
EXAONE + SFT
22.9 (+12.5)
59.4 (+6.7)
66.2 (+1.3)
Table 1. Performance across out-of-distribution generalization benchmarks, and in-distribution genre coverage.
OOD (Generalization)
ID (Coverage)
Data
Arena Hard
WritingBench
Multi-Genre
DeepWriting
4.9
55.5
65.9
LongWriter
7.0
56.4
69.6
Multi-Genre
7.5
59.1
71.4
Table 2. Comparison with existing writing data using 2K sampled subsets, with Qwen3-8B fine-tuned on each corpus.
# Genres
0
4
8
12
All
Novelty ( ↑ )
3.87
3.93
4.26
4.56
4.81
Table 3. Effect of genre count on creative output novelty. More genres yield more varied, non-redundant outputs.
Judge
GPT-5-mini
DeepSeek v3.2
Gemini 3 Flash
Llama
43.7
36.6
41.7
Llama + SFT
60.6
63.1
70.5
Qwen
56.1
51.5
58.5
Qwen + SFT
63.6
65.9
71.5
Table 4. WritingBench evaluation under independent judges.