The arXiv preprint reports that LiteraryBigFive performed better than the listed comparison methods when used to generate author-personalized text from four held-out books. The reported advantage appeared across ROUGE, SIM, GPT-4 and human evaluations.
A common map for literary style
The method represents authorial writing in a unified, interpretable five-dimensional space. Its five axes are Classicism, Ornateness, Narrativity, Emotionality and Analyticity. The axes are built from semantics-preserving contrasts between an authorial passage and a neutral version.
For a target book, the method estimates authorial coordinates from reference passages and updates the model’s internal activations toward those coordinates during generation. This provides a route to personalized steering through a book-specific position in the shared space.
During decomposition, the method removes a shared principal component described as overall expressiveness. Mean absolute off-diagonal cosine similarity, a measure of how closely the directions align, was 0.87 before decomposition and 0.27 afterward.
The benchmark
Passages were segmented into blocks from 120 to 400 tokens. The axis-construction corpus contained 1,322 passages and 12,741 sentences, while the held-out evaluation set contained 590 passage samples and 5,716 sentences.
The main experiments used Llama2-7B-Chat. The comparison included few-shot prompting, LLM-Steer, LoRA, ICV, Mean-Centering, CAA and RepE. An additional set of experiments used Qwen2.5-3B-Instruct, where the authors report best performance across all four books and all evaluation metrics.
The reported scores
Across the aggregate results, LiteraryBigFive scored 81.6 on GPT-4 Semantic Fidelity and 57.7 on GPT-4 Authorial Adherence. The corresponding human ratings were 84.6 and 59.3. These were the highest reported aggregate scores.
An ablation compared the full model with versions without decomposition or without the style gap. With the measures ordered as ROUGE-1, ROUGE-L, SIM and GPT-4, the full model scored 53.0, 44.0, 95.2 and 69.6. The version without decomposition scored 52.1, 43.2, 94.7 and 68.8, while the version without the style gap scored 49.3, 40.0, 91.6 and 66.6. In this comparison, both component-removed versions had reported scores below the full model on all four measures.
Signals behind the model
The estimated coordinates had a mean Pearson correlation of r=0.96 with an ensemble of style judgments from GPT-5, Claude 3.5 and Gemini 3 across all five axes. The result is correlational and relies on LLM judgments rather than an independent human consensus.
Layer-wise probing achieved AUC above 0.90 across all five stylistic dimensions, indicating that information about each dimension was linearly decodable in the tested model. Human annotators achieved Cohen’s kappa of 0.59, described as moderate inter-annotator agreement, and the authors state that GPT-4 scores closely aligned with human evaluations.
The reported complexity was O(K · d), with K=6 intervention vectors. Average latency was 19.88 milliseconds per token over 100 generated cases, compared with 18.93 for Mean-Centering and 19.92 for LLM-Steer.
A narrow test
The authors report that the axes are primarily derived from English literary classics, steering acts globally on residual-stream layers, and the method requires access to internal activations, limiting its use with closed-source APIs. The document is an arXiv preprint, and journal publication or peer-review status is not reported.
Paper data and sources
Original title: LITERARYBIGFIVE: Author-Personalized Text Generation in a Unified Interpretable Space
Authors: Jinghui Zhang, Lang Gao, Ao Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text