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

It introduces AuthorMix, a modular authorship style transfer framework that decomposes style modeling into reusable style-specific LoRA adapters rather than covering all target styles with a single model at once. The abstract states that existing methods 'train a single model on large corpora to model all target styles at once,' which is high-cost and offers limited flexibility for target-specific adaptation; AuthorMix replaces this monolithic approach with modular adapters. The evidence comes from the abstract-level method description and baseline comparison statements; concrete experimental settings, data scale, and statistical details are not given in the loaded text.

Through reinforcement-learning-based layer-wise adapter mixing, a specialized adaptation model can be trained rapidly for each new target author using only a handful of target-style training examples. Relative to monolithic methods that require large-corpus training, this mechanism lowers the adaptation cost for a new target to the order of 'a handful of target-style training examples.' The abstract explicitly states the 'handful of target-style training examples' requirement, but the loaded text reports no specific sample counts or ablation results.

It ranks first on the combined style-meaning score, surpassing all baselines including GPT-5.1, and substantially improves meaning preservation over the trained baselines. The abstract frames this as jointly addressing style transfer and meaning preservation, whereas existing methods 'often sacrifice meaning preservation for style transfer.' This is an automatic-evaluation conclusion reported in the abstract; metric definitions, the baseline set, and numerical values are not expanded in the loaded text.

Under human evaluation, AuthorMix is the only method that is best-or-tied on every dimension. This statement extends the advantage from a single combined score to multi-dimensional human judgment, though the abstract does not say on which dimensions other methods fall behind. This is a human-evaluation conclusion reported in the abstract; the evaluation dimensions, number of raters, and agreement information are not given in the loaded text.

The work targets the authorship style transfer setting: rewriting text in a target author's style while preserving the original meaning. Its design intent is that adapting to a new target author requires only a handful of target-style examples, making it better suited to settings with many target authors and limited samples per target; the advantages described in the abstract were obtained under its own baseline set and evaluation setup, so applicability to other languages, genres, or longer texts needs separate confirmation for a given use case.

The loaded text is the abstract and metadata page of the paper, without the body, figures, tables, or appendix, so the exact composition of the combined style-meaning score, the full list of baselines, the actual number of target-style training examples, and the dimensions and procedure of the human evaluation cannot be verified. In addition, the abstract does not describe performance on low-resource authors, cross-lingual settings, or long texts, nor whether the layer-wise mixing weights carry an interpretable structural meaning; these are points to watch in further reading and verification.