Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction

• URL: http://arxiv.org/abs/2605.27878v1
• Date: Wed, 27 May 2026 02:55:34 GMT
• Title: Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction
• Abstract summary: We construct a matched story-continuation paradigm across StoryStar, TMAS, and The New Yorker.<n>We compare continuations from four OLMo 32B checkpoints against matched human text.<n>Because these checkpoints share architecture, scale, tokenizer, and pretraining, the design isolates the post-training effect.
• Score: 3.8920533503590007
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models produce fluent fiction, yet their creative output is widely seen as flat. We ask where this quality originates in the training and whether it affects different domains of human fiction equally. We construct a matched story-continuation paradigm across StoryStar (public-platform), TMAS (prompt-guided), and The New Yorker (professional literary)-and compare continuations from four OLMo 32B checkpoints (Base, SFT, DPO, RLVR) against matched human text. Because these checkpoints share architecture, scale, tokenizer, and pretraining, the design isolates the post-training effect. We measure each continuation along three sentence-level dimensions: thematic motion, affective prevalence, and linguistic diversity. Across all three, post-training compresses dynamic variation: thematic transitions become more uniform, high-intensity emotions give way to neutrality, and stylistic diversity across stories shrinks. We term this progressive loss narrative flattening. The effect is directionally stable across story domains but gap size depends on the human baseline: professional literary fiction is compressed most, while public-platform and prompt-guided stories show smaller gaps, consistent with their human baselines sitting closer to the model's default rhythm. Post-trained endpoints converge across domains, suggesting alignment produces a continuation regime largely insensitive to the source domain's narrative texture.

Related papers

• Do Large Language Models Always Tell The Same Stories? [2.4779770815686444]
  We investigate the diversity of large language models-generated stories through the framework of narrative similarity.<n>Our findings reveal a consistent trend: LLM-generated narratives are consistently more similar to each other than human-written stories are.<n>We show that common mitigation strategies, including negative prompting and temperature scaling, fail to meaningfully address this homogeneity.
  arXiv  Detail & Related papers  (2026-06-15T22:52:02Z)
• StoryScope: Investigating idiosyncrasies in AI fiction [26.913725703912572]
  We find that AI-generated stories cluster in a shared region of narrative space, while human-authored stories exhibit greater diversity.<n>More broadly, these results suggest that differences in underlying narrative construction, not just writing style, can be used to separate human-written original works from AI-generated fiction.
  arXiv  Detail & Related papers  (2026-04-03T15:56:38Z)
• Generating Visual Stories with Grounded and Coreferent Characters [63.07511918366848]
  We present the first model capable of predicting visual stories with consistently grounded and coreferent character mentions.<n>Our model is finetuned on a new dataset which we build on top of the widely used VIST benchmark.<n>We also propose new evaluation metrics to measure the richness of characters and coreference in stories.
  arXiv  Detail & Related papers  (2024-09-20T14:56:33Z)
• Are Large Language Models Capable of Generating Human-Level Narratives? [114.34140090869175]
  This paper investigates the capability of LLMs in storytelling, focusing on narrative development and plot progression. We introduce a novel computational framework to analyze narratives through three discourse-level aspects. We show that explicit integration of discourse features can enhance storytelling, as is demonstrated by over 40% improvement in neural storytelling.
  arXiv  Detail & Related papers  (2024-07-18T08:02:49Z)
• Measuring Psychological Depth in Language Models [50.48914935872879]
  We introduce the Psychological Depth Scale (PDS), a novel framework rooted in literary theory that measures an LLM's ability to produce authentic and narratively complex stories. We empirically validate our framework by showing that humans can consistently evaluate stories based on PDS (0.72 Krippendorff's alpha) Surprisingly, GPT-4 stories either surpassed or were statistically indistinguishable from highly-rated human-written stories sourced from Reddit.
  arXiv  Detail & Related papers  (2024-06-18T14:51:54Z)
• Dynamic Typography: Bringing Text to Life via Video Diffusion Prior [73.72522617586593]
  We present an automated text animation scheme, termed "Dynamic Typography" It deforms letters to convey semantic meaning and infuses them with vibrant movements based on user prompts. Our technique harnesses vector graphics representations and an end-to-end optimization-based framework.
  arXiv  Detail & Related papers  (2024-04-17T17:59:55Z)
• CPST: Comprehension-Preserving Style Transfer for Multi-Modal Narratives [1.320904960556043]
  Among static visual narratives such as comics and manga, there are distinct visual styles in terms of presentation. The layout of both text and media elements is also significant in terms of narrative communication. We introduce the notion of comprehension-preserving style transfer (CPST) in such multi-modal domains.
  arXiv  Detail & Related papers  (2023-12-14T07:26:18Z)
• DeltaScore: Fine-Grained Story Evaluation with Perturbations [69.33536214124878]
  We introduce DELTASCORE, a novel methodology that employs perturbation techniques for the evaluation of nuanced story aspects. Our central proposition posits that the extent to which a story excels in a specific aspect (e.g., fluency) correlates with the magnitude of its susceptibility to particular perturbations. We measure the quality of an aspect by calculating the likelihood difference between pre- and post-perturbation states using pre-trained language models.
  arXiv  Detail & Related papers  (2023-03-15T23:45:54Z)
• Inferring the Reader: Guiding Automated Story Generation with Commonsense Reasoning [12.264880519328353]
  We introduce Commonsense-inference Augmented neural StoryTelling (CAST), a framework for introducing commonsense reasoning into the generation process. We find that our CAST method produces significantly more coherent, on-topic, enjoyable and fluent stories than existing models in both the single-character and two-character settings.
  arXiv  Detail & Related papers  (2021-05-04T06:40:33Z)
• PlotMachines: Outline-Conditioned Generation with Dynamic Plot State Tracking [128.76063992147016]
  We present PlotMachines, a neural narrative model that learns to transform an outline into a coherent story by tracking the dynamic plot states. In addition, we enrich PlotMachines with high-level discourse structure so that the model can learn different writing styles corresponding to different parts of the narrative.
  arXiv  Detail & Related papers  (2020-04-30T17:16:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.