StoryLens: Preference-Aligned Story Rewriting via Context-Aware Narrative Enrichment

Abstract: Story rewriting aims to adapt existing narratives to diverse reader preferences while preserving plot consistency and narrative coherence. Unlike conventional work on style transfer, we argue that effective story rewriting demands context-aware narrative enrichment beyond surface-level stylistic adaptation. Our pilot human study shows that style adaptation alone provides only marginal gains in reader satisfaction (2.3%), while context-enhanced rewriting substantially improves user preference alignment (24.5%). Motivated by this, we introduce STORYLENSBENCH, a large-scale benchmark for preference-aligned story rewriting, comprising structured story books, multi-dimensional reader preference profiles, and ranked context-aware rewritten stories. Building on this benchmark, we propose STORYLENSEVAL, a reward model for estimating reader satisfaction over rewritten stories, and STORYLENSWRITER, a two-stage rewriting model combining supervised fine-tuning with GRPO-based reinforcement learning. We further establish a comprehensive evaluation framework covering fidelity, coherence, and reader satisfaction. Experimental results demonstrate that STORYLENSWRITER consistently outperforms strong generation and personalization baselines, highlighting the importance of context-aware narrative enrichment for personalized story rewriting.

• The paper introduces StoryLens, a context-aware rewriting framework that enriches descriptions, emotions, and character interactions while preserving plot fidelity and reader preferences.
• StoryLensBench uses 163 novels, 780 Goodreads reader profiles, five preference dimensions, and human-verified rewrite rankings, while StoryLensEval reaches 0.8947 human correlation and 85.87% accuracy.
• StoryLensWriter combines supervised fine-tuning with GRPO and a length-control reward, raising preference win rate to 0.84, PerSE to 7.78, coherence to 93.11, and local fidelity to 94.62.

Motivation and problem formulation

Story rewriting differs from open-ended story generation in that a model must preserve an existing narrative structure while adapting the text to reader-specific preferences. The authors argue that this task cannot be reduced to style transfer: satisfying diverse readers requires context-aware narrative enrichment—expanding scene descriptions, refining emotional expression, deepening character interactions—while remaining anchored to the established plot. They formalize the task as generating $y^{*} = \arg\max_{y} P_{\theta}(y \mid x, c, p)$, where $x$ is the original segment, $c$ is structured context from preceding chapters, and $p$ is a multi-dimensional preference profile.

The central empirical claim is established by a pilot study with 10 participants over 8 books: style adaptation alone raises satisfaction only marginally (3.43 → 3.51), whereas context-enhanced rewriting lifts satisfaction to 4.27 with only minor fidelity loss (4.62 → 4.41). This 24.5% relative improvement in preference alignment versus 2.3% for style adaptation alone motivates the entire framework.

StoryLensBench

The benchmark is built through a four-step pipeline: chapter segmentation of 163 novels from BookSum and Project Gutenberg across seven genres; extraction of structured narrative contexts (writing background, characters, events, relations) from preceding storylines; preference modeling from real Goodreads reviews of 780 active readers, distilled by Gemini 3.1 Pro into a five-dimensional taxonomy (Plot, Characters, Language, Worldview, Theme); and generation plus ranking of five variant rewrites per chapter per profile, converted into pairwise comparisons stratified into easy/medium/hard difficulty levels. Human verification on sampled cases yields an 81% inter-annotator ranking agreement. Unlike synthetic persona datasets, the use of authentic review histories grounds preferences in realistic reader behavior; the keyword analysis shows a long-tail distribution across all five dimensions, indicating that rewriting must capture both dominant trends and fine-grained individual signals.

StoryLensEval

StoryLensEval is a Qwen3-8B backbone with an MLP scoring head trained via LoRA using a pairwise softplus ranking objective $$\mathcal{L}_{sat} = \mathrm{softplus}(-(s_w - s_l))$$. Training follows an easy-to-hard curriculum followed by score-head refinement on hard pairs. Validation against human annotations (40 pairs, majority voting among three annotators) shows correlations of 0.8947 with humans and 0.9250 with strong LLM judges, substantially above comparably scaled LLM judges such as Qwen3-8B (0.4211). Ablations confirm both design choices: a dual-tower variant drops overall accuracy from 85.87 to 73.40, indicating that joint encoding of profile and text matters for fine-grained preference interactions, and removing the second refinement stage costs roughly 10 points overall.

StoryLensWriter

StoryLensWriter uses Qwen3.5-9B as its backbone with a two-stage pipeline. Stage one performs LoRA-based SFT on the paired rewriting data to instill task format and basic preference-aware behavior. Stage two applies GRPO on approximately 5% of the SFT prompts, sampling groups of candidate rewrites scored by StoryLensEval as reward, augmented with a length-control term—a Huber-style penalty keeping rewrites within 0.7–1.3× the original length—with total reward $R = R_{\mathrm{rm}} + 0.15 R_{\mathrm{length}}$.

The GRPO stage produces a stark contrast: relative to the SFT-only model, win rate rises from 20.44 to 79.56, PerSE improves from 7.59 to 7.78, and coherence increases from 91.10 to 93.11. This indicates that the learned evaluator provides a usable optimization signal rather than merely a measurement instrument.

Evaluation framework

The evaluation protocol covers three dimensions. Local fidelity combines SVO triplet matching and Narrative Entity Knowledge Graph preservation (weighted 0.1/0.9 toward NEKG), with GPT-4o arbitration for fuzzy embedding-similarity cases. Global fidelity extracts claims and labels them Entailment, Neutral, or Contradiction against both the original text and the structured context, explicitly treating globally plausible literary expansions as acceptable rather than hallucination—an important design decision given that enrichment is the point of the task. Coherence follows BooookScore's eight error types at sentence level. Satisfaction is measured externally via PerSE under the Per-MPST setting and internally via StoryLensEval win/top-1 rates, with the latter restricted to benchmarking open-source baselines to mitigate self-evaluation bias from the LLM-driven data construction.

Main results

Against five strong prompted LLMs (Gemini-3.1-Pro, GPT-5.2, DeepSeek-V3, Doubao-Seed-2.0-Lite, Qwen-VL-Max), StoryLensWriter achieves the best PerSE (7.78) along with the highest local fidelity (94.62) and global fidelity (98.98), while remaining competitive on coherence and NovelCritique. Against comparably sized baselines, the margins are larger:

Method	Local Fid.	Global Fid.	Coh.	Sat. Win	Sat. Top-1	PerSE
Llama-3.1-8B	92.54	96.96	80.53	0.09	0.00	7.52
Qwen3.5-9B	94.12	98.64	90.72	0.52	0.03	7.68
Hybrid RAG (Qwen3.5-9B)	93.97	96.26	87.54	0.64	0.18	7.50
PPlug	94.35	99.71	80.99	0.70	0.30	7.65
StoryLensWriter	94.62	98.98	93.11	0.84	0.47	7.78

The comparison supports a clear conclusion: retrieval of historical contexts or injection of persona embeddings is insufficient for long-form rewriting; explicit two-stage optimization yields substantially better preference alignment (win rate 0.84 vs. 0.70 for the strongest personalization baseline) without sacrificing narrative quality. The case study on Robinson Crusoe illustrates the mechanism: context enhancement ties mercantile preference cues back to Crusoe's established background, producing continuity-grounded characterization rather than decorative stylistic matching.

Limitations and open questions

The paper concedes several constraints. The benchmark is English-only and derives preferences exclusively from Goodreads, so multilingual, cross-cultural, and contemporary reading preferences remain untested. Source texts are limited to BookSum and Project Gutenberg literary narratives; generality to web novels, children's literature, interactive stories, or domain-specific narratives is unexamined. Methodologically, the satisfaction metric partly relies on StoryLensEval itself, whose training data was produced by Gemini 3.1 Pro—the authors mitigate but do not eliminate this circularity by restricting self-evaluation to open-source baselines and validating against external PerSE scores. Whether the evaluator's rankings generalize beyond the five-dimension taxonomy and the seven genres represented in training is left open.

Conclusion

This work reframes personalized story rewriting as context-aware narrative enrichment rather than style transfer, substantiated by pilot evidence that contextual enrichment drives nearly all achievable gains in reader satisfaction. It contributes a benchmark grounded in real reader reviews with ranked rewrites, a validated pairwise preference evaluator, and a two-stage SFT-plus-GRPO writer that outperforms both frontier prompted LLMs and dedicated personalization methods on preference alignment while preserving fidelity and coherence. The results collectively indicate that input-stage personalization mechanisms are inadequate for long-form narrative adaptation, and that a trained preference reward model can serve effectively as both evaluator and RL signal.