The paper introduces a training‑free method for uncovering prompt‑conditional stylistic axes in large language models (LLMs). By repeatedly sampling completions of a single prompt at high temperature and applying Principal Component Analysis (PCA) to the pooled hidden activations, the authors automatically label the resulting axes using the extreme (pole) generations. Validation against 245 human‑elicited stylistic annotations shows that, for the Qwen‑3.5‑4B‑Instruct model, the top two axes align with human dimensions with 72.8% precision and 43.6% macro‑recall, and 75.6% of validity ratings confirm the axes’ polar generations, while other models exhibit varying degrees of discoverability.
By Ajit Mallavarapu, Ziwei Gu
The paper investigates whether activation steering can capture the complex, multidimensional nature of authorship style. By using structured contrastive prompting along rhetorically motivated dimensions, the authors construct rich style representations directly in activation space, revealing a shared authorship backbone with aspect‑specific residuals. They propose Aspect‑Aware Activation Steering (A3S), a training‑free framework that merges per‑aspect contrastive directions, employs interference‑aware aggregation, and tunes steering strength per instance, achieving better authorship style transfer and outperforming a trained baseline on out‑of‑domain benchmarks.
By Hieu Tran, Calvin Bao, Marine Carpuat
AuthorMix is a lightweight, modular framework for authorship style transfer that uses layer‑wise adapter mixing. It trains individual style‑specific LoRA adapters on a small set of high‑resource authors, enabling rapid adaptation to new target styles with only a few examples. The method achieves the highest combined style‑meaning score among baselines, including GPT‑5.1, and improves meaning preservation, as confirmed by human evaluation.
By Sarubi Thillainathan, Ji-Ung Lee, Michael Sullivan, Alexander Koller
HyperStyler is a new architecture for low-resource authorship style transfer that separates style selection from style realization. It uses a style navigator to predict style coordinates from source context and target-author references, and a style hypernetwork to apply these coordinates through dynamic parameter modulation. Experiments on Reddit, Blog, and News datasets show that HyperStyler outperforms previous methods, including LLM-based approaches, while adding only 2.4% more parameters than T5-large and running 1.8× faster at inference.
By Jongkyung Shin, Minguk Jeon, Chanwoo Park, Chiehyeon Lim
The paper "Evaluating Style-Personalized Text Generation: Challenges and Directions" examines the difficulties of assessing text that is tailored to individual users’ styles. It critiques common metrics such as BLEU, embeddings, and LLM-as-judges, and introduces a style discrimination benchmark covering domain discrimination, authorship attribution, and LLM-generated personalized versus non-personalized discrimination across eight writing tasks. The study finds that ensembles of diverse evaluation metrics outperform single-evaluator approaches and offers guidance for reliable assessment of style-personalized generation.
By Anubhav Jangra, Bahareh Sarrafzadeh, Silviu Cucerzan, Adrian de Wynter, Sujay Kumar Jauhar
RT‑SFT is a method for text style transfer that uses roundtrip translation through a pivot language to strip stylistic information from monolingual corpora, creating pseudo‑parallel data. This data is then used to LoRA‑finetune an instruction‑tuned large language model as a stylizer, allowing the model to rewrite sentences in a target style while preserving meaning. Experiments across four style domains show that RT‑SFT surpasses state‑of‑the‑art approaches, including few‑shot in‑context learning, and offers effective retrieval augmentation for expert style domains with strict terminology.
By Ruoxi Liu, Philipp Koehn
arXiv:2507. 18043v2 Announce Type: replace-cross Abstract: Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights.
By Duy Nguyen, Archiki Prasad, Elias Stengel-Eskin, Mohit Bansal
arXiv:2609.07037v1 Announce Type: new
Abstract: Activation steering has emerged as a lightweight, inference-time approach to control the behavior of Large Language Models (LLMs). However, traditional...
By Takeru Hiramatsu, Kyohei Atarashi, Koh Takeuchi, Hisashi Kashima
arXiv:2606.18389v2 Announce Type: replace
Abstract: Large language models (LLMs) have become an effective tool for synthetic data generation, including for low-resource languages, where generated dat...
By Jan Cegin, Daniil Gurgurov, Yusser Al Ghussin, Simon Ostermann
Authorship verification (AV) assumes that an author's writing style remains sufficiently stable to distinguish it from that of other writers. In practice, however, this assumption is challenged by dis...
arXiv:2606. 10099v1 Announce Type: cross Abstract: The rapid development of large language models (LLMs) has raised concerns about misuse such as plagiarism, misinformation, and automated influence operations, motivating the need for robust detectors.
By Rafael Rivera Soto, Barry Chen, Nicholas Andrews
arXiv:2510.13302v4 Announce Type: replace-cross
Abstract: Computational stylometry studies writing style through quantitative textual patterns, enabling applications such as authorship attribution, i...
By Pablo Miralles-Gonz\'alez, Javier Huertas-Tato, Alejandro Mart\'in, David Camacho