LLM Fiction Suffers 'Narrative Flattening' Due to Post-Training

A new research paper titled "Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction" investigates the phenomenon of large language models producing creative writing that is perceived as lacking depth. The study compares LLM outputs from various training stages (Base, SFT, DPO, RLVR) against human-written fiction across different domains. Findings indicate that post-training processes significantly reduce thematic transitions, emotional intensity, and stylistic diversity in LLM-generated stories, a effect termed 'narrative flattening'. This flattening is observed across all story types but is most pronounced in professional literary fiction, suggesting that alignment techniques may make LLM outputs less sensitive to the original narrative style. AI

IMPACT This research highlights how current LLM training methods may be limiting creative expression, suggesting a need for new approaches to preserve stylistic and thematic richness in AI-generated fiction.

RANK_REASON The cluster contains a research paper detailing findings on LLM creative writing capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction
• OLMo 32B
• StoryStar
• The New Yorker

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →