AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing

Read full text Summarise with VIRAT
Authors: Sarubi Thillainathan, Ji-Ung Lee, Michael W. Sullivan, Alexander Koller
Cited by: 0
References: 32

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The task of authorship style transfer involves rewriting text in the style of a target author while preserving the meaning of the original text.Exist ing style transfer methods train a single model on large corpora to model all target styles at once: this highcost approach offers limited flex ibility for targetspecific adaptation, and often sacrifices meaning preservation for style transfer.In this paper, we propose AuthorMix: a light weight, modular, and interpretable style transfer framework.We first train individual, stylespecific LoRA adapters on a small set of highresource authors: this allows for the rapid training of spe cialized adaptation models for each new target using layerwise adapter mixing via reinforcement learning, necessitating only a handful of target style training examples.AuthorMix ranks first on the combined style-meaning score among all baselines, including GPT5.1, and substantially improves meaning preservation over the trained baselines; under human evaluation it is the only method bestortied on every dimension.1 Approach MIS Output Source 1.000 "I want to tell you something," I said: "I want to tell you all."ASTRAPOP 0.271 I want to tell you,' he said.TinyStyler 0.800 I wanted to tell you all, and tell you all, and tell you all, and tell you all, and tell you all, and tell you all.STYLL 0.731 I have a burning desire to share with you everything, to lay it all out on the table."GPT 5.116shot 0.768 I broke in at last."I have something to say," I began; "I would lay it bare before every one of you."AuthorMix 0.905 "Listen, I have something to say to you," I declared."I have a message for all of you," I exclaimed.Neutral 0.830 The speaker expressed a desire to convey information, stating that they wished to disclose everything.STYLL ASTRAPOP TinyStyler AuthorMix Method Fewshot prompting Condition on fewshot SFT + DPO/CPO Condition on Embedding SFT + selfdistill Adapter mixing

Fields of Study & Subjects

Authorship Attribution and Profiling Generative Adversarial Networks and Image Synthesis Topic Modeling

Publication details

Journal / source
Conference on Empirical Methods in Natural Language Processing (EMNLP)
Pages
50382–50416
Publication date
1 January 2026
Language
EN