Paper deep dive
AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing
Sarubi Thillainathan, Ji-Ung Lee, Michael Sullivan, Alexander Koller
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/26/2026, 1:44:10 AM
Summary
AuthorMix is a modular, lightweight authorship style transfer framework that uses layer-wise adapter mixing to adapt pre-trained LoRA adapters to new target styles. By optimizing mixing weights via reinforcement learning (GRPO) or gradient-free methods (LoRAHub) on a small set of target examples, it achieves superior meaning preservation and style transfer performance compared to existing baselines like GPT-5.1 and ASTRAPOP, particularly in low-resource settings.
Entities (5)
Relation Signals (3)
AuthorMix → uses → LoRA
confidence 100% · We train individual, style-specific LoRA adapters on a small set of high-resource authors
AuthorMix → outperforms → GPT-5.1
confidence 95% · AuthorMix outperforms existing, SoTA style-transfer baselines -- as well as GPT-5.1
GRPO → optimizes → AuthorMix
confidence 90% · We additionally consider directly learning W using the Group Relative Policy Optimization (GRPO...)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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. Existing style transfer methods train a single model on large corpora to model all target styles at once: this high-cost approach offers limited flexibility for target-specific adaptation, and often sacrifices meaning preservation for style transfer. In this paper, we propose AuthorMix: a lightweight, modular, and interpretable style transfer framework. We train individual, style-specific LoRA adapters on a small set of high-resource authors, allowing the rapid training of specialized adaptation models for each new target via learned, layer-wise adapter mixing, using only a handful of target style training examples. AuthorMix outperforms existing, SoTA style-transfer baselines -- as well as GPT-5.1 -- for low-resource targets, achieving the highest overall score and substantially improving meaning preservation.
Tags
Links
- Source: https://arxiv.org/abs/2603.23069v1
- Canonical: https://arxiv.org/abs/2603.23069v1
Trouble viewing inline? Open PDF directly →
Full Text
63,511 characters extracted from source content.
Expand or collapse full text
AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing Sarubi Thillainathan Ji-Ung Lee Michael Sullivan Alexander Koller Saarland University sarubi|msullivan|koller@lst.uni-saarland.de, ji-ung.lee@uni-saarland.de 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 flexibility for targetspecific adaptation, and often sacrifices meaning preservation for style transfer. In this pa per, we propose AuthorMix: a lightweight, mod ular, and interpretable style transfer framework. We train individual, stylespecific LoRA adapters on a small set of highresource authors, allowing the rapid training of specialized adaptation mod els for each new target via learned, layerwise adapter mixing, using only a handful of target style training examples. AuthorMix outperforms existing, SoTA styletransfer baselines—as well as GPT5.1—for lowresource targets, achieving the highest overall score and substantially improv ing meaning preservation. 1 1 Introduction Generating text that matches a targeted writing style is crucial for many applications such as per sonalized tutors, language learning, and conversa tional agents, among others (Xu et al., 2025). An important aspect of fostering user engagement is properly adjusting the style of the generated text. This is typically referred to as authorship style transfer and framed as a rewriting task in which a given text is rewritten to match the style of a specific target author (Patel et al., 2024; Liu et al., 2024a). While style transfer methods have versatile use cases, they face two major challenges: (i) the rewritten text must preserve the original meaning, while (i) the style must change substantially 1 Code and data will be made publicly available. Figure 1: Overview of AuthorMix. (1) We train indi vidual LoRA adapters for each highresource author. (2) Scalar mixing weights are optimized to jointly max imize target style and meaning preservation. (3) The resulting model rewrites any text into the target style. enough to match the target style. These challenges are further compounded in lowresource settings where examples of the target style are limited. To address these challenges, many methods first rewrite the source text into a neutral style and then rewrite it into the target style (Liu et al., 2024a; Patel et al., 2024). Other approaches avoid the need for a neutral pivot style by adjusting the style directly in the embedding space, utilizing pre trained style embeddings (Horvitz et al., 2024b; Riley et al., 2021; Horvitz et al., 2024a). All exist ing methods involve on training a single, general ApproachMISOutput Original 1.000“I want to tell you something,” I said: “I want to tell you all.” ASTRAPOP 0.271I want to tell you,‘ he said. TinyStyler 0.800I 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 0.731I have a burning desire to share with you everything, to lay it all out on the table.“ GPT 5.1 16shot 0.768I broke in at last. “I have something to say,” I began; “I would lay it bare before every one of you.” AuthorMix (Ours) 0.905“Listen, I have something to say to you,” I declared. “I have a message for all of you,” I exclaimed. Neutral 0.830The speaker expressed a desire to convey information, stating that they wished to disclose everything. Table 1: Outputs of different styletransfer methods for a single sentence along with their mutual implication score (MIS) that captures how well the meaning is preserved. See analysis in Appendix D.5.1 purpose model—an approach that often struggles to adapt to new rewrite styles (see Table 1). In contrast, we propose AuthorMix, a light weight, modular approach involving two stages of training. In the first stage, we train individual LoRA adapters (Hu et al., 2022) for a few high resource authors (styles). For each target style, we then construct a model by learning (layerwise) mixing weights over these adapters using a small set of targetstyle examples. In comparison to existing styletransfer meth ods that require extensive pretraining or inter mediate neutralstyle text, AuthorMix directly produces a model tailored to a specific target style under a lightweight training regimen (see Figure 1). Moreover, our layerwise weight mixing improves interpretability by revealing which high- resource authors contribute at which layer. Finally, our weight mixing objective explicitly balances style transfer and meaning preservation, encourag ing the modelgenerated rewrites’ semantic faith fulness to the source text (see Table 2). Experiments demonstrate that AuthorMix out performs current SoTA rewrite models—and GPT5.1—at meaning preservation, while match ing the current SoTA for style transfer in low resource settings. We further find that layerwise mixing with weights learnt via reinforcement learning (RL) substantially outperforms static adapterwise mixing; identifying layerspecific, gradientbased mixing as a promising direction for future work in model merging. Our contributions are as follows: 1.AuthorMix, a lightweight, modular, and interpretable authorship style transfer frame work capable of efficiently producing a style transfer model for any given target author. 2.Extensive evaluations comparing Au thorMix to current SoTA style transfer Source: ‘I will,’ he said; and instantly went off through a gate, Lizzy continuing her way. Output 1: ‘I will,’ he muttered, and darted off through the gate, while Lizzy continued on her way. Output 2: “Agreed,” he exclaimed, and at once strode through the gate, while Lizzy continued on her path. Output 3: He replied, “I will,” and instantly went off through a gate, while Lizzy continued on her way. Output 4: He replied, then swiftly exited through a gate, while Lizzy continued on her path. Output 5: “Yes, I will,” he said, and hurried through the gate, while Lizzy went on her way. Table 2: AuthorMix rewrites a single source sentence into different target author styles. Each output is pro duced by the same base model with different pertarget mixing weights. models on 100 source–target author pairs, demonstrating that AuthorMix more effec tively preserves semantic content in low resource settings than existing approaches without sacrificing styletransfer capability. 2 Related Work Jin et al. (2022) broadly categorize style transfer methods into (pseudo)parallel and nonparallel data methods. Whereas early parallel data works heavily rely on large parallel corpora (Xu et al., 2012) (Krishna et al., 2020), later works increas ingly focus on lowresource scenarios with only a few available target style texts. The lack of large parallel corpora is often mit igated via intermediate neutral style texts (i.e., pseudoparallel data) which are then used as in context learning examples (Patel et al., 2024) or even for policy optimization (Liu et al., 2024a). While neutral style texts compensate for a lack of parallel data, using them can lead to a semantic drift and result in lower meaning preservation; STYLLASTRAPOPTinyStylerAuthorMix Method Few shot prompting Condition on few shot SFT + DPO/CPO Condition on Embedding SFT + selfdistill Adapter mixing Trainset authors –All 10 highresource authors ∼1M∼4 authors Neutralization YesYesNoNo Adapt to new target Change prompt Change prompt Change embedding Optimize weights Modular / decomposable NoNoNoYes Interpretable NoNoNoYes (perlayer) Training Cost –48 h> 48 h ∼8.7 h Table 3: Comparison of authorship style transfer approaches. See Appendix D.4 for Training cost details. especially when they are also required during in ference (and not only during training). To avoid neutralization, nonparallel data methods often operate in a style embedding space, aiming to directly adjust the style (RiveraSoto et al., 2021) (Wegmann et al., 2022). One recent work that utilizes style embeddings is TinyStyler, that relies upon a multistage training pipeline to remove the reliance on neutral style text during inference (Horvitz et al., 2024b). Two common shortcomings of existing methods is that they rely upon costly training (using thou sands of author pairs or millions of instances) and that they train a single monolithic model to serve all target styles, making it difficult to extend to new target styles. AuthorMix alleviates these issues by providing a lightweight means to produce a spe cific target style model (see Table 3). 3 Methodology We separate authorship style transfer into three steps, which induces modularity, eases extensibil ity, and allows us produce a targetstyle specific model (see Figure 1): 1. Author Adapter Training: We train indi vidual style adapters for a few high resource authors where abundant text is available. 2. (Layer-wise) Weight Mixing: For a given target author, we use a few text examples to conduct (layerwise) weight mixing of indi vidual style adapters. 3. Text Rewriting: We use the resulting model during inference to rewrite any source text provided by a source author. 2 2 Note, that the target and source author might be the same (i.e., the user) in various realworld scenarios. However, this distinction is important for evaluation (see Section 4). 3.1 Author Adapter Training We first train individual LoRA adapters (Hu et al., 2022) to rewrite source texts into the style 3 of a specific, highresource target author 푎. Given a corpus 풳︀ 푎 of texts from author 푎, we construct pseudoparallel training pairs by paraphrasing each 푥∈풳︀ 푎 into a neutral version ̃푥—removing vocabulary choices, sentence rhythm, syntactic patterns—while preserving the underlying content and meaning, by prompting a Language Model (see Appendix B.1). We then finetune a separate LoRA adapter to reconstruct the original text 푥 from the styleremoved version ̃푥: ℒ︀ SFT =− 1 |풳︀ 푎 | ∑ 푥∈풳︀ 푎 1 |푥| ∑ |푥| 푖=1 log푝 휃+휃 푎 (푥 푖 |̃푥,푥 :푖 ) where 휃 are the frozen base model parameters and 휃 푎 are the LoRA parameters for author 푎. Given 푛 high resource authors 푎 1 ,...,푎 푛 , we train the set 풜︀=휃 1 ,...,휃 푛 of corresponding authorspecific adapters. 3.2 Weight Mixing In our second step, we conduct a weighted linear mixing of the trained adapters 휃 (푖) ∈풜︀ into a sin gle targetstyle adapter ̂ 휃. While similar mixing ap proaches have been explored before with adapter specific weights (Huang et al., 2024; Ilharco et al., 2023; Fisher et al., 2024), to the best of our knowledge we are the first to conduct a more fine grained mixing using layerspecific weights. For a model with 퐿 layers, the 푗 푡ℎ layer ̂ 휃 푗 (1≤푗≤ 퐿) of the targetstyle adapter ̂ 휃 is then the linear combination of weights 푊 푖,푗 across 푛 adapters: 3 By style, we refer to recurring characteristics of how an author writes, such as lexical choice, syntactic preferences, punctuation, formatting, and other surfacelevel linguistic patterns, rather than the underlying semantic content. ̂ 휃 푗 =∑ 푛 푖=1 푊 푖,푗 ⋅휃 (푖) 푗 We use 푊∈ℝ 푛×퐿 to refer to the matrix of mixing weights. Despite the computational overhead com pared to an adapterwise mixing, we conjecture that the greater expressiveness will help in accom modating authorspecific characteristics that may be located at different layers. This is in line with findings in model interpretability research that attribute specific linguistic knowledge to different layers (Tenney et al., 2019; Fayyaz et al., 2021; Zhou and Srikumar, 2022). Given a lowresource target style (author) 푡 with a small set 풳︀ 푡 of example texts written by 푡, we ob tain 푊 by optimizing with respect to a styletrans fer specific objective function 푆(푥 푠 ,푥 푠→푡 ,풳︀ 푡 ) (the Joint Score of Section 4.5). For each 푥 푠 in our set 풳︀ 푠 of training examples, and each corresponding candidate rewrite 푥 푠→푡 of 푥 푠 , 푆(푥 푠 ,푥 푠→푡 ,풳︀ 푡 ) is defined as the geomet ric mean of: (i) the Toward Score (Patel et al., 2024) 푇(푥 푠 ,푥 푠→푡 ,풳︀ 푡 ) between the style embed ding of 푥 푠→푡 and the mean style embedding of 풳︀ 푡 , which measures the stylistic similarity between the candidate rewrite and the target author’s exam ple texts; and (i) the Mutual Implication Score (Babakov et al., 2022) MIS(푥 푠 ,푥 푠→푡 ), which mea sures the semantic similarity between the original sentence 푥 푠 and the rewrite 푥 푠→푡 (see Section 4.5). 푆(푥 푠 ,푥 푠→푡 ,풳︀ 푡 )= √ 푇(푥 푠 ,푥 푠→푡 ,풳︀ 푡 )×푀퐼푆(푥 푠 ,푥 푠→푡 ) In both of our optimization approaches (Sec tion 3.2.1 and Section 3.2.2), the base model and all LoRA parameters remain frozen, and only the mixing weights 푊 are updated. 3.2.1 LoRAHub In our first approach, following the LoRAHub framework (Huang et al., 2024), we use Nevergrad’s AutoOptimizer (NGOpt, Rapin and Teytaud (2018)). This metaoptimizer automat ically selects an appropriate gradientfree opti mization strategy for the given problem. Under this approach, the model 휋 푊 (푖) generates a candidate rewrite 푥 푠→푡 for each 푥 푠 ∈풳︀ 푠 at each iteration 푖. The weights 푊 (푖+1) for the next itera tion are then optimized via NGOpt to maximize the objective ℒ︀ LH : ℒ︀ LH =∑ 푥 푠 ∈풳︀ 푠 푆(푥 푠 ,푥 푠→푡 ,풳︀ 푡 )+휆⋅|푊 (푖) | Following Huang et al. (2024), we incorporate the L1 regularization term 휆⋅|푊 (푖) | to mitigate overly large weights. 3.2.2 GRPO We additionally consider directly learning 푊 using the Group Relative Policy Optimization (GRPO; Shao et al. (2024)) reinforcement learning (RL) algorithm, with our joint score 푆(푥 푠 ,푥 푠→푡 ,풳︀ 푡 ) serving as the reward function. To the best of our knowledge, this is the first application of RLbased policy optimization to learning adapterbased mixing weights. Prior work on adapter mixing has relied exclusively on gradientfree methods (Huang et al., 2024; Fisher et al., 2024) or fixed heuristics (Ilharco et al., 2023; Wortsman et al., 2022). 3.3 Adapter Subset Selection While weight mixing can be conducted using all adapters available in 풜︀, this does not scale well with an increasing number of adapters, especially for gradientfree mixing, which often relies on searchbased algorithms. Moreover, considering irrelevant adapters for weight mixing can even hurt the performance of the resulting model (see our ablation study in Section 5.3). To address this, we restrict to the subset of 풜︀ corresponding to the top푘 most stylistically similar authors to the target style for weight mixing (Fisher et al., 2024). We quantify style similarity using the cosine dis tance between prototypical style embeddings 푒 푎 for highresource author 푎 and target style embed dings 푒 푡 . We compute 푒 푎 by randomly sampling a set of texts 풳︀ ′ 푎 ⊂풳︀ 푎 and averaging the resulting embeddings: 푒 푎 = 1 |풳︀ ′ 푎 | ∑ |풳︀ ′ 푎 | 푖=1 Emb(푥 푖 ) We similarly obtain a representative embedding for the target style 푡: 푒 푡 = 1 |풳︀ 푡 | ∑ |풳︀ 푡 | 푖=1 Emb(푥 푖 ) Finally, we select the top푘 most similar adapters based on the resulting ranking using the cosine distance cos(푒 푡 ,푒 푎 ) between all available authors and the target style. 4 Experimental Setup Our experiments include six baselines as well as four different variations of AuthorMix (adapter and layerwise mixing using LoRAHub and GRPO methods). We report averaged results across ten runs using different seeds for pseudorandom number generation. 4.1 Datasets We utilize three disjoint sets of authors with Eng lish literary text from Project Gutenberg (Gerlach and FontClos, 2018): • High-resource authors. Liu et al. (2024b) provide a collection of 10 highresource au thors with substantial amounts of available text (more than 2,000 texts each). We use these to train our individual adapter models (see Section 3.1). • Target authors. We randomly sample 10 unique authors from Project Gutenberg with 16 target texts each. This mimicks a low resource scenario where only a small amount of text is available for a given target style. • Source authors. Finally, we randomly sam ple another 10 unique authors from Project Gutenberg as source authors, whose texts serve as input to be styletransferred (풳︀ 푠 in Section 3.2). For each source author, we randomly split their texts into a training set of 50 texts for RLbased weight mixing (Sec tion 3.2.2) and reserve 16 texts for testing. Following Patel et al. (2024) and Liu et al. (2024a), we sample 16 test examples for each sourcetarget pair, resulting in 10×10=100 evaluation pairs and a total of 1,600 test instances. We provide detailed statistics for all 30 authors in Appendix C.1. 4.2 Baselines We compare AuthorMix against various base lines including promptingbased, embeddingcon ditioned, and policyoptimization methods. For all baselines, we use the same hyperparameters as reported in their respective original papers unless otherwise noted. Table 3 summarizes the key dif ferences between different style transfer methods. • Neutral Text Baseline. This serves as a lower bound and simply rewrites source texts into a neutral style using LLaMA3.370B Instruct. • Few-shot Prompting. We prompt GPT 5.1 to perform style transfer and provide 16 target style texts as incontext examples alongside the source text to be transferred. • STYLL. Patel et al. (2024) propose an in context learning approach using pseudopar allel texts. We replicate their work using 16 incontext examples and respective style de scriptors (see Appendix C.2 for more details). • TinyStyler. Horvitz et al. (2024b) propose a method that utilizes style embeddings dur ing training and is capable of style transfer without requiring pseudoparallel text during inference. We directly use their benchmarked released checkpoint. To ensure a fair compar ison against singleinstance baselines, we do not generate multiple candidates with a re ranking. • ASTRAPOP. Liu et al. (2024a) investigate supervised finetuning (SFT) as well as direct preference optimization (DPO) to train mod els for style transfer. Due to the lack of pub licallyavailable model weights, we retrain their approach with a same amount of data (20k samples). Specifically, we used 2,000 texts from each of the 10 highresource au thors and trained ASTRAPOP with LLaMA 3.1 8B Instruct as the base model. We report results for both the SFTonly variant and the full, DPOoptimized model. 4.3 AuthorMix Variants AuthorMix conducts a targetspecific weight mix ing either via LoRAHub; see Section 3.2.1 or GRPO; see Section 3.2.2. To better assess the impact of each weight mixing variant as well as the impact of our adapter subset selection (Sec tion 3.3), we evaluate four variants: • LoRAHub (adapter-wise). The adapter wise weight mixing method proposed by Huang et al. (2024). • LoRAHub (layer-wise). The LoRAHub framework extended to layerwise mixing. • GRPO (adapter-wise). We utilize GRPO to learn single mixing weights for each adapter. • GRPO (layer-wise). We utilize GRPO to learn layerwise mixing weights for each adapter. We further evaluate all AuthorMix configurations for different 푘∈2,...,10 of most stylistically similar authors. 4.4 Base Model and Hyperparameters We use LLaMA3.18BInstruct (Touvron and others, 2023) as the base model for AuthorMix. All highresource author adapters are imple mented as LoRA (Hu et al., 2022) modules and finetuned using the LLaMAFactory framework (Zheng et al., 2024). The neutral paraphrases that are used for adapter training are generated using a LLaMA3.370BInstruct model with zero shot prompting (Appendix B.1). All experiments were conducted on a highperformance computing cluster with 4×NVIDIA H100/A100 GPUs. 4 We provide detailed hyperparameters in Appen dix C.3. 4.5 Evaluation Metrics Following Patel et al. (2024), we report three pri mary metrics: • Toward. Measures how far the transferred output has moved toward the target author’s style, as a fraction of the maximum possible movement. Given source text 푥 푠 , target author 푡, and transferred text 푥 푠→푡 with respective style embeddings 푒 푠 , 푒 푡 , and 푒 푠→푡 , the toward score is computed as: Toward= max(푆푖푚(푒 푠→푡 ,푒 푡 )−푆푖푚(푒 푠 ,푒 푡 ),0) 1−푆푖푚(푒 푠 ,푒 푡 ) , where 푆푖푚 is the angular simi larity 푆푖푚(푢,푣)=1−arccos(푢⋅푣/(‖푢‖⋅ ‖푣‖))/휋 over style embeddings from STAR (HuertasTato et al., 2023) — a RoBERTa large encoder pretrained on 4.5M texts from 70K authors (including authors from Project Gutenberg). 5 • Semantic preservation (MIS). The mutual implication score (MIS, Babakov et al. (2022)) evaluates bidirectional entailment be tween the source and transferred texts using an NLI model. MIS ranges between 0 and 1 and has been shown to correlate well with human judgments of meaning preservation in style transfer tasks (Babakov et al., 2022; Patel et al., 2024). 4 The CPU architectures varied depending on the used node. 5 LUARMUD (RiveraSoto et al., 2021) was not suitable for our selected author set; see Gaafar et al. (2025) for discus sion. Method Toward ↑MIS ↑Joint ↑ Neutral Text† 0.010.790.05 Few shot (GPT5.1)0.080.810.20 STYLL0.070.680.16 ASTRAPOPSFT0.160.630.29 ASTRAPOPDPO 0.17 0.630.29 TinyStyler0.160.750.31 AuthorMix LoRAHub, AW (푘=3) 0.110.850.25 LoRAHub, LW (푘=8) 0.12 0.87 0.27 GRPO, AW (푘=7) 0.130.780.29 GRPO, LW (푘=4) 0.160.83 0.34 Table 4: Main results averaged over 100 source–target author pairs. ↑ = higher is better. Bold = best, under lined = second best. †Reference bound (neutralized text, no style transfer). AW = adapterwise mixing; LW = layerwise mixing. Secondary metrics (CoLA, Away) are reported in Table 7. • Joint. The geometric mean of Toward and MIS: Joint= √ Toward×MIS, capturing the tradeoff between style transfer and mean ing preservation. We also report Away (Patel et al., 2024) and CoLA (Warstadt et al., 2019) as secondary evaluation metrics in Appendix C.4: Away is more relevant to sourcestyle obfuscation than to targetstyle trans fer, while CoLA mainly serves as a fluency safe guard for adapter mixing. AuthorMix maintains high fluency (CoLA = 0.83–0.87), and its Away scores are on par with other methods (0.640.73). 5 Results & Analysis Table 4 reports results for AuthorMix and all base lines, averaged across all 100 source–target author pairs. We report the bestperforming 푘 (for adapter subset selection) for each AuthorMix configura tion and provide the full results across all 푘 in Figure 4 (detailed results are provided in Table 7). The results show that AuthorMix achieves the best joint score (0.34) with layerwise GRPO weight mixing (푘=4), followed by TinyStyler (0.31) and ASTRAPOPDPO (0.29). With respect to level of granularity for the weight mixing, we find that a finegrained, layerwise mixing consis tently outperforms an adapterwise weight mixing. 0.0 0.2 0.4 0.6 0.8 1.0 MIS Score AuthorMixASTRAPOP-SFTASTRAPOP-DPOTinyStylerSTYLLFewShotNeutral Figure 2: Distribution of MIS scores across all 100 source–target pairs for each method. Boxes show interquartile range (IQR); diamonds mark the mean; whiskers extend to 1.5×IQR. The dashed line at 0.5 indicates the threshold below which semantic equiva lence breaks down. 5.1 Other Metrics Both AuthorMix and TinyStyler operate directly on source text without a neutralization step, yet AuthorMix achieves a substantially higher MIS (0.83 vs. 0.75) while matching TinyStyler’s Toward score (both 0.16). Methods that do uti lize neutral text during inference (ASTRAPOP, STYLL) achieve comparable or higher Toward scores, but at the cost of lower MIS (∼ 0.63). While the MIS differences may appear mod est numerically, Table 1 illustrates how scores around 0.75 already introduce noticeable meaning changes. Figure 2 shows the distribution of MIS scores across all instances and random seeds. AuthorMix has a consistently high MIS score with low variance. In contrast, all author style transfer methods either have a lower mean and median MIS and/or a larger variation. We provide detailed distributions and methodwise variances in Appendix D.2. With respect to fluency, AuthorMix variants achieve consistently high CoLA scores of 0.83– 0.87, comparable to all other baselines except both TinyStyler variants (which scores 0.68). This is also reflected in the TinyStyler example in Table 1. The low fluency of TinyStyler might be attributed to its informal Reddit text training data (Horvitz et al., 2024b) (in contrast to other methods trained on Project Gutenberg data) and the resulting do main mismatch might further contribute to the fluency gap when applying this model to literary text. Overall, our results suggest that AuthorMix with a modular, pertarget weight mixing achieves a competitive target style transfer with a substan tially higher meaning preservation compared to methods that train large, monolithic models. Figure 3: Global average of learned layer weights across all 10 target authors and all 푘 values for GRPO layerwise mixing. Red bars indicate positive weights (adapter contribution is used); blue bars indicate neg ative weights (adapter is suppressed). Vertical lines indicate ±1 standard deviation across target authors. A pertarget author heatmap is provided in Figure 8. 5.2 LoRAHub vs. GRPO A key design choice in AuthorMix is the selec tion of the weight mixing method. Comparing GRPO against LoraHub under identical conditions (same 푘) and similar compute budget reveals that across all configurations, GRPO achieves higher Toward scores; with larger differences for layer wise mixing (vs. authorwise mixing). Although LoRAHub variants achieve a higher MIS (0.85– 0.87 vs. 0.78–0.83), they tradeoff suboptimally with the Toward score (as reflected in their lower joint scores). We can further see that GRPO ben efits substantially more compared to LoRAHub from a low number 푘 of highresource author adapters (Figure 4). Finally, we analyze the individually learned weights in each layer, and visualize the averaged weights for GRPO and LoRAHub separately. Fig ure 3 shows the global average of learned layer weights across all 10 target authors and all 푘 con figurations (푘=2–10) for GRPO layerwise mixing. Interestingly, we find that GRPO learns structured layerwise patterns, with the middle layers (14– 21) receiving the strongest positive weights, while early layers (0–10) show small mixedsign weights and deep layers (24–31) are predominantly nega tive, suppressing adapter contributions at these levels. The high standard deviation across target authors reflects that different target authors require different layer compositions, reinforcing the need for pertarget optimization. A pertarget author breakdown is provided in Figure 8. In contrast, LoRAHub assigns nearzero weights across most layers (see Appendix D.3.1), indicating that it is not suited well for a layerwise mixing. We con jecture that the substantially larger search space 012345678910 Number of adapters k 0.15 0.20 0.25 0.30 0.35 Joint score Toward × MIS k=0 (base LLM) k=1 (single) LoRAHub, adapter-wise LoRAHub, layerwise GRPO, adapter-wise GRPO, layerwise Cost (h/pair) 0 5 10 15 20 25 Cost (h / pair) Figure 4: Joint score (left axis) and wallclock training cost in hours per target (right axis) vs. number of related authors 푘 for all four AuthorMix variants. 푘=0 is the base LLaMA3.18BInstruct (zeroshot, no adapter); 푘=1 is a single bestmatching adapter. All scores averaged over 100 source–target pairs. (compared to adapterwise mixing) results in an optimization problem that is too difficult to solve well using the gradientfree methods available in NGOpt. We leave investigating a mixedinteger linear programing approach using more sophisti cated solvers such as Gurobi (Gurobi Optimization LLC, 2024) for future work. 5.3 Impact of Number of Adapters 푘 We further conduct an ablation study to better understand the impact of using different 푘 high re source author adapters. Figure 4 shows the results for all four AuthorMix variants. We can see that all scores sharply rise from 푘=1 (single adapter, ≈ 0.17 joint score) to 푘=2 (0.23–0.32), confirming that mixing even two adapters is substantially bet ter than using a single one. We further observe that GRPO layerwise peaks at 푘=4 (Joint = 0.34) and remains competitive through 푘=6 before degrad ing at 푘≥7. In contrast, LoRAHub variants plateau earlier around 푘=3 and remain consistent for in creasing 푘. This aligns with findings by Fisher et al. (2024) who report that mixing more than five adapters hurts grammaticality and findings by Huang et al. (2024) who observe that mixing more adapters expands the search space, leading to a higher variance. We conclude that identifying a good subset of top푘 stylistically similar authors is essential as it reduces the search space and ensures that each adapter contributes a meaningful stylistic signal. 5.4 Analyzing Resource Consumption We further provide an analysis on the resource consumption for each methods, highlighting the efficiency of AuthorMix (see Table 3). For in stance, on comparable compute infrastructure, ASTRAPOP requires ∼48h of LoRA SFT + DPO training on all 10 highresource authors (Liu et al., 2024a). TinyStyler is pretrained using ∼8M texts from over 1M users (Horvitz et al., 2024b). In con trast, AuthorMix selects a small, targeted subset of authors and adapts to any new target by learning a total of only ∼128 scalar mixing weights which takes ∼40 minutes on a single GPU and requires no model retraining. Moreover, new adapters can be added independently to 풜︀ with LoRA SFT requiring ∼2h of training on a single GPU. AuthorMix provides an easily extensible alter native to large, monolithic models where produc ing a target style specific model incurs a cost of ∼ 40 min. Our results indicate that this pertarget in vestment leads to a genuine pertarget optimization —maximizing style transfer and meaning preser vation for each target style. 6 Conclusion We introduced AuthorMix, a modular framework for lowresource authorship style transfer that composes a target style specific LoRA adapter via optimized mixing weights. Unlike prior methods that train a single, monolithic model on large cor pora, AuthorMix builds a modular adapter library from a small number of highresource authors and dynamically adapts them to any new target. Our analysis shows that our novel GRPObased method for learning a weight mixing results in interpretable, layerwise weights; revealing that most stylistically important information is cen tered around the middle layers. In future work, we plan to explore more sophis ticated solvers for gradientfree weight mixing (as the ones implemented in LoRAHub failed to find meaningful weights). Moreover, the improved results for layerwise mixing (over adapterwise mixing) indicate that there is a high untapped potential which would also intrinsically improve interpretability (using the mixing weights). We believe that studying more advanced model merg ing strategies such as TIESMerging (Yadav et al., 2023) or learned gating mechanisms may yield further improvements and are directly compatible with our framework. Since AuthorMix is modular, it naturally extends to crosstask style transfer by combining style adapters with taskspecific ones (e.g., for summarization or dialogue) to produce stylized outputs for tasks beyond paraphrasing, and to scale larger, more diverse adapter libraries. 7 Limitations Our evaluation is conducted on a single domain (Project Gutenberg literary texts) and a single base model (LLaMA3.18BInstruct). While the framework is architectureagnostic, we have not yet demonstrated generalization to other base models (e.g., Qwen) or to domains with substan tially different stylistic characteristics such as informal social media text or technical writing. All authors in our dataset are historical nonfiction/ fiction writers from a similar time period; the method’s effectiveness for authors with highly distinctive or contemporary styles remains to be validated. The adapter library currently contains 10 high resource authors. Scaling to a larger and more diverse library could improve coverage of a wider range of target styles, but it would also increase the onetime cost of adapter training (∼2hrs). How ever, this cost is incurred only once, and the same highresource adapters can then be reused across multiple target authors. In return, AuthorMix pro vides a more interpretable approach to authorship style transfer. In addition, although AuthorMix requires only about 40 minutes of weight learning for each new target, this pertarget adaptation cost is not zero, unlike in baselines that adapt through prompt or embedding changes at no additional cost. Inference cost, however, remains unchanged across all AuthorMix variants, regardless of 푘, be cause the mixed adapters are merged into the base model; it is therefore comparable to that of base lines built on the same LLaMA3.18BInstruct backbone. Finally, our evaluation relies on automatic met rics (MIS, CoLA, Toward). The STAR model used for evaluation also guides adapter selection and weight optimization, mirroring the design of StyleRemix (Fisher et al., 2024), which likewise uses its style scores for both axis selection and optimization. Weight learning uses sourceauthor texts disjoint from the test set. While MIS corre lates well with human judgments of meaning preservation (Babakov et al., 2022), a human eval uation and validation with an independent style metric would further strengthen the findings. 8 Ethical Considerations Authorship style transfer presents important ethi cal risks, as models that reproduce a person’s writing style could be used for impersonation, deception, or other forms of misuse. In this work, we study the task in the context of publicly avail able literary texts and position it as a research problem aimed at understanding and improving personalized text adaptation, rather than enabling unauthorized imitation. Nevertheless, improved style transfer quality may also increase misuse potential. We therefore emphasize that any down stream deployment should include safeguards such as transparency about generated content, restric tions on modeling private or living individuals without consent, and careful consideration of po tential harms related to identity, authorship, and misuse. References Nikolay Babakov, David Dale, Varvara Logacheva, and Alexander Panchenko. 2022. A LargeScale Compu tational Study of Content Preservation Measures for Text Style Transfer and Paraphrase Generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 300–321. Alexandra Chronopoulou, Matthew E. Peters, Alexan der Fraser, and Jesse Dodge. 2023. AdapterSoup: Weight Averaging to Improve Generalization of Pre trained Language Models. In Findings of the Associ- ation for Computational Linguistics: EACL 2023, pages 2009–2025. Mohsen Fayyaz, Ehsan Aghazadeh, Ali Modarressi, Hosein Mohebbi, and Mohammad Taher Pilehvar. 2021. Not All Models Localize Linguistic Knowl edge in the Same Place: A Layerwise Probing on BERToids’ Representations. In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 375– 388, Punta Cana, Dominican Republic. Jillian Fisher, Skyler Hallinan, Ximing Lu, Mitchell L Gordon, Zaid Harchaoui, and Yejin Choi. 2024. StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Moemen Gaafar, Anurag Sarkar, and Matthew Guzdial. 2025. Fewshot StyleConditioned LLM Text Gener ation via Latent Interpolation. Martin Gerlach and Francesc FontClos. 2018. A Stan dardized Project Gutenberg Corpus for Statistical Analysis of Natural Language and Quantitative Lin guistics. Entropy 22. Gurobi Optimization LLC. 2024. Gurobi Optimizer Reference Manual. Carolin Holtermann, Markus Frohmann, Navid Rekab saz, and Anne Lauscher. 2024. What the Weight?! A Unified Framework for ZeroShot Knowledge Com position. In Findings of the Association for Compu- tational Linguistics: EACL 2024, pages 1138–1157, St. Julian’s, Malta. Zachary Horvitz, Ajay Patel, Chris CallisonBurch, Zhou Yu, and Kathleen McKeown. 2024a. Paraguide: Guided diffusion paraphrasers for plug andplay textual style transfer. In Proceedings of the AAAI conference on artificial intelligence, pages 18216–18224. Zachary Horvitz, Ajay Patel, Kanishk Singh, Chris CallisonBurch, Kathleen McKeown, and Zhou Yu. 2024b. TinyStyler: Efficient FewShot Text Style Transfer with Authorship Embeddings. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13376–13390. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan AllenZhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: LowRank Adaptation of Large Language Models. In International Confer- ence on Learning Representations (ICLR). Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. 2024. LoraHub: Effi cient CrossTask Generalization via Dynamic LoRA Composition. In First Conference on Language Mod- eling. Javier HuertasTato, Alejandro Martin, and David Ca macho. 2023. Understanding Writing Style in Social Media with a Supervised Contrastively Pretrained Transformer. arXiv preprint arXiv:2310.11081. Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Worts man, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. Editing Models with Task Arithmetic. In International Conference on Learning Represen- tations (ICLR). Ganesh Jawahar, Benoît Sagot, and Djamé Seddah. 2019. What Does BERT Look At? An Analysis of BERT’s Attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP. Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea. 2022. Deep Learning for Text Style Transfer: A Survey. Computational Linguistics 48(1):155–205. Kalpesh Krishna, John Wieting, and Mohit Iyyer. 2020. Reformulating Unsupervised Style Transfer as Para phrase Generation. In Proceedings of the 2020 Con- ference on Empirical Methods in Natural Language Processing, pages 737–762. Shuai Liu, Shantanu Agarwal, and Jonathan May. 2024a. Authorship style transfer with policy opti mization. arXiv preprint arXiv:2403.08043. Xinyue Liu, Harshita Diddee, and Daphne Ippolito. 2024b. Customizing Large Language Model Gener ation Style using ParameterEfficient Finetuning. In Proceedings of the 17th International Natural Language Generation Conference, pages 412–426, Tokyo, Japan. Ajay Patel, Nicholas Andrews, and Chris Callison Burch. 2024. LowResource Authorship Style Trans fer: Can NonFamous Authors Be Imitated? In Proceedings of the AAAI Conference on Artificial Intelligence. Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: NonDestructive Task Composition for Transfer Learning. In Proceedings of the 16th Conference of the European Chapter of the Associ- ation for Computational Linguistics (EACL), pages 487–503. Jeremy Rapin and Olivier Teytaud. 2018. Nevergrad — A GradientFree Optimization Platform. Parker Riley, Noah Constant, Mandy Guo, Girish Kumar, David Uthus, and Zarana Parekh. 2021. TextSETTR: FewShot Text Style Extraction and Tunable Targeted Restyling. In Proceedings of the 59th Annual Meeting of the Association for Compu- tational Linguistics and the 11th International Joint Conference on Natural Language Processing (Vol- ume 1: Long Papers), pages 3786–3800, Online. Rafael A. RiveraSoto, Olivia Elizabeth Miano, Juanita Ordonez, Barry Y. Chen, Aleem Khan, Marcus Bishop, and Nicholas Andrews. 2021. Learning Uni versal Authorship Representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 913–919, On line and Punta Cana, Dominican Republic. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv preprint arXiv:2402.03300. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT Rediscovers the Classical NLP Pipeline. In Proceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 4593– 4601, Florence, Italy. Hugo Touvron and others. 2023. Llama 2: Open Foun dation and FineTuned Chat Models. arXiv preprint arXiv:2307.09288. Weizhong Wang, Chao Du, Binhang Yuan, Zirui Liu, and Pengcheng He. 2024. LoRA Soups: Merging LoRAs for Practical Skill Composition Tasks. arXiv preprint arXiv:2410.13025. Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural Network Acceptability Judg ments. Transactions of the Association for Computa- tional Linguistics 7:625–641. Anna Wegmann, Marijn Schraagen, and Dong Nguyen. 2022. Same author or just same topic? towards contentindependent style representations. arXiv preprint arXiv:2204.04907. Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael GontijoLopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. 2022. Model Soups: Averaging Weights of Multiple FineTuned Models Improves Accuracy without Increasing Inference Time. In International Conference on Machine Learning, pages 23965– 23998, PMLR. Wei Xu, Alan Ritter, Bill Dolan, Ralph Grishman, and Colin Cherry. 2012. Paraphrasing for Style. In Proceedings of COLING 2012, pages 2899–2914, Mumbai, India. Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and TatSeng Chua. 2025. Personalized Gener ation In Large Model Era: A Survey. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 24607–24649. Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. 2023. TIESMerging: Resolving Interference When Merging Models. In Advances in Neural Information Processing Systems (NeurIPS). Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, and Yongqiang Ma. 2024. Lla maFactory: Unified Efficient FineTuning of 100+ Language Models. In Proceedings of the 62nd An- nual Meeting of the Association for Computational Linguistics: System Demonstrations. Yichu Zhou and Vivek Srikumar. 2022. A Closer Look at How Finetuning Changes BERT. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1046–1061, Dublin, Ireland. ARelated Work A.1Adapter Merging and Composition LoRA (Hu et al., 2022) enables parameterefficient finetuning, and composing multiple adapters has become an active research area. AdapterFusion (Pfeiffer et al., 2021) introduces a twostage approach that first trains taskspecific adapters independently, then learns to compose them via a learned attention mechanism — separating knowledge extraction from knowledge composi tion. AdapterSoup (Chronopoulou et al., 2023), (Holtermann et al., 2024) show that weightaver aging domainspecific adapters, selected by simi larity to the test input, improves outofdomain generalization without any additional training. Lo RAHub (Huang et al., 2024) introduces gradient free optimization over a library of taskspecific LoRA modules, learning scalar mixing weights from a few examples to achieve competitive cross task generalization. StyleRemix (Fisher et al., 2024) applies a similar composition idea to text style: it trains LoRA modules on distinct stylistic axes (formality, sentence length, lexical complex ity) and combines them with optimized adapter level weights to steer outputs toward desired style dimensions for authorship obfuscation. More re cently, LoRA Soups (Wang et al., 2024) shows that optimally weighted concatenation of LoRAs outperforms naive averaging on skill composition tasks. Our work builds directly on this paradigm: rather than composing task or styleaxis adapters, we compose author-style adapters and optimize pertarget mixing weights to steer generation to ward a specific individual’s writing style. BMethodology B.1Pseudo-Parallel Data Generation For each highresource author’s text, we gener ate a neutral paraphrase using LLaMA3.370B Instruct with the prompt: “Paraphrase the fol- lowing text into a neutral, encyclopedic tone, removing stylistic markers while keeping meaning unchanged.” The resulting (neutral, original) pairs serve as supervised finetuning data for the author adapters. CExperimental Setup C.1Author Lists Project Gutenberg is an open repository of literary works by authors whose writings have entered the public domain 6 . Its eBooks may be used for a wide range of purposes, including the creation of deriv ative works, reports, performances, and research. All texts used in this work were obtained from Project Gutenberg and are drawn from English language authors. High-resource (adapter) authors. Mark Twain, Virginia Woolf, Vernon Lee, Charlotte Perkins Gilman, George Orwell, Jane Austen, Nathaniel Hawthorne, Oscar Wilde, P. G. Wode house, and Samuel Richardson (obtained from (Liu et al., 2024b) 7 ). Low-resource target authors. Arthur Conan Doyle, Joseph Conrad, Jack London, E. M. Forster, F. Scott Fitzgerald, Rudyard Kipling, Robert W. Chambers, Agatha Christie, Edith Wharton, and L. M. Montgomery. Source authors. L. Frank Baum, Charlotte Brontë, Lewis Carroll, Charles Dickens, Thomas Hardy, G. A. Henty, John Ruskin, Walter Scott, Henry B. Wheatley, and David Widger. C.2Baseline Details We evaluate STYLL on GPT3.5turbo since the GPT3 endpoint used in (Patel et al., 2024) is deprecated by OpenAI, and GPT3.5turbo is the closest available model. We did not include BLOOM7B based on STYLL, since MIS is in low range. C.3Base Model and Hyperparameters C.3.1LoRA Adapter Configuration All adapters are finetuned using LLaMAFactory (Zheng et al., 2024) with the following configura tion: LoRA rank 8, 훼=16, applied to all linear target modules, input cutoff length of 256 tokens, and maximum output length of 256 tokens. high- resource author adapters and AuthorMix trained using following template: Listing 1 (no other ad ditional input like previous works). For weight optimization, merging, and inference we largely follow the hyperparameter choices of StyleRemix (Fisher et al., 2024) (e.g., top푝=0.95, tempera ture 휏=1.0). C.3.2Weight Learning Details For each target author, we learn the layerwise mixing weights using GRPO (Section 3.2.2) on the heldout sourceauthor texts: each candidate 6 https://w.gutenberg.org/policy/license.html 7 https://github.com/cauchy221/StyleTunedLM "instruction": "Paraphrase", "input": "By 2020 and beyond this number may have increased even further.", "output": "it is possible that this number will increase even later by 2020." Listing 1: Example inputoutput pair used in adapter finetuning for AuthorMix. HyperparameterValue OptimizerNGOpt Optimization steps250 Weight initialization0 Weight bounds[−1.5 to +1.5] Top 푝 sampling 0.95 Temperature 휏 1.0 seed42 Table 5: LoRAHub, Gradientfree optimization hyper parameters. weight configuration generates a batch of style transferred texts, which are scored against the target author’s reference texts using the joint re ward (Section 3.2.1). This is conceptually similar to the policy optimization stage of (Liu et al., 2024a) however, in our case the base model and all adapter parameters are frozen — only the 푚×퐿 scalar mixing weights are tuned (e.g., 4× 32=128 weights with 푘=4 related authors and LLaMA3.18B’s 32 transformer layers). For each input, the model samples a group of 퐺 candidate outputs, and the mixing weights are updated using relative reward rankings within each group. C.3.3Gradient-Free Optimization Hyperpara- meters For the gradientfree learning (Section 4.3), we use NGOpt (Rapin and Teytaud, 2018) as provided by LoraHub (Huang et al., 2024) with the config urations provided in Table 5. We use the compa rable compute budgets to match with GRPO based weight learning ( 40mints). C.3.4GRPO Hyperparameters We use hyper parameters as given in Table 6. we set 훽 controls the optional KL regularization to zero. HyperparameterValue Learning rate0.02 Optimization steps300 Top 푝 sampling 0.95 Temperature 휏 1.0 Weight initialization0 Random seeds42 Table 6: GRPO optimization hyperparameters. C.4On Away and CoLA as Secondary Metrics Away score. The Away metric measures how far the transferred text has moved away from the source author’s style, expressed as a fraction of the source–target distance (Patel et al., 2024). We ex clude it from the Joint score for two reasons. First, Away is underdetermined: many stylistically dif ferent outputs can achieve the same Away value, since any text that is equidistant from the source in the embedding space scores identically — regard less of whether it moved toward the target or in an unrelated direction. This makes Away a weak sig nal for evaluating style transfer quality. Second, a high Away score primarily matters in authorship obfuscation settings — such as those targeted by StyleRemix (Fisher et al., 2024) — where the explicit goal is to hide the source author’s identity. In standard style transfer, the objective is to approach the target style (captured by Toward) while preserving content (captured by MIS); a text can succeed on both without fully erasing source style traces, which is acceptable in most practical applications. CoLA score. Large language mod els, especially instructiontuned variants like LLaMA3.18BInstruct, generally produce fluent and grammatical text, making fluency a less discriminative metric in recent work. However, (Fisher et al., 2024) (who trained on Base LLaMa model) observe that adapter merging — which is central to our approach — can introduce fluency degradation, particularly when merging five or more adapters. Since AuthorMix composes multi ple LoRA adapters via learned weights, we include CoLA as a safeguard to verify that the merging process does not degrade linguistic acceptability. Our results confirm that AuthorMix maintains high CoLA scores (0.83–0.87), comparable to Method 푘Toward ↑MIS ↑Joint ↑CoLA ↑Away ↑ Neutral Text† —0.010.790.05 0.890.91 Zeroshot (LLaMA8B)—0.070.810.170.880.81 Few shot (GPT5.1)—0.080.810.200.720.65 STYLL (GPT3.5)—0.070.680.160.880.86 STYLL (GPT5.1)—0.150.620.270.710.71 ASTRAPOPSFT—0.160.630.290.800.63 ASTRAPOPDPO— 0.17 0.630.290.830.78 TinyStyler—0.160.750.310.680.66 TinyStylerSim—0.160.800.330.680.63 Single adapter (푘=1) 10.060.860.170.730.46 AuthorMix — LoRAHub, adapter-wise 20.100.840.230.860.74 30.110.850.250.850.71 40.100.840.240.850.73 50.110.840.250.840.70 60.110.850.250.830.67 70.110.860.240.810.61 80.100.840.240.840.69 90.110.860.250.820.62 100.100.850.240.810.63 AuthorMix — LoRAHub, layerwise 20.110.860.260.830.65 30.120.860.270.830.64 40.120.860.270.830.65 50.120.860.260.830.65 60.12 0.87 0.270.830.64 70.120.860.270.830.64 80.12 0.87 0.270.830.64 90.12 0.87 0.270.830.64 100.12 0.87 0.270.830.64 AuthorMix — GRPO, adapter-wise 20.120.790.270.870.75 30.130.780.290.860.72 40.120.770.260.850.73 50.110.750.250.860.77 60.130.780.280.880.75 70.130.780.290.870.73 80.120.760.270.860.75 90.120.760.270.860.76 100.120.730.260.860.77 AuthorMix — GRPO, layerwise 20.150.830.320.850.69 30.140.830.310.850.71 40.160.83 0.34 0.830.67 50.150.820.320.840.72 60.150.820.330.830.71 70.110.800.260.830.77 80.120.820.280.840.76 90.130.800.300.830.77 100.110.790.260.820.80 Table 7: Full results for all baselines and all AuthorMix configurations across 푘∈2,...,10 related authors. All scores averaged over 100 source–target pairs. Bold = best within each AuthorMix variant. †Reference bound. nonmerged baselines, indicating that the learned weight composition does not harm fluency. AuthorMix ASTRAPOP-SFT ASTRAPOP-DPO TinyStyler TinyStyler-Sim STYLL FewShot Neutral 0.0 0.2 0.4 0.6 0.8 1.0 MIS Score Figure 5: Box plot of MIS score distributions across all 100 source–target pairs for each method (including TinyStylerSim). Boxes show the interquartile range; diamonds mark the mean; whiskers extend to 1.5×IQR; circles are outliers. The dashed line at 0.5 indicates the threshold below which semantic equivalence breaks down. DResults & Analysis D.1Full Results Table Table 7 reports the complete results including TinyStylerSim and all AuthorMix configurations across 푘∈1,...,10. D.2MIS Score Distributions Figure 5 and Figure 6 show the full MIS score distributions across all 100 source–target pairs, in cluding TinyStylerSim postpublication variant. The KDE plot (Figure 6) annotates each method with its mean (휇) and standard deviation (휎). AuthorMix achieves the highest mean MIS (휇=0.83) with the lowest variance (휎=0.059) among all styletransfer methods — nearly half the standard deviation of ASTRAPOPSFT (휎= 0.115) and ASTRAPOPDPO (휎=0.107). This confirms that AuthorMix not only preserves mean ing better on average but does so consistently across source–target pairs. TinyStylerSim (휇=0.80, 휎=0.072) im proves over TinyStyler (휇=0.75, 휎=0.071) by adding further candidate generation and filtering candidates for semantic similarity, narrowing the gap to AuthorMix. However, AuthorMix still achieves a +4% higher mean with lower variance, and without requiring any candidate filtering step. ASTRAPOP variants show the widest spread, with density mass extending well below the 0.5 threshold (dashed line in Figure 5), confirming that aggressive style transfer via DPO/SFT fre quently comes at the cost of meaning. STYLL (휇=0.68, 휎=0.099) occupies a middle ground, while FewShot (휇=0.81) and Neutral (휇=0.79) score high on MIS but apply minimal style trans fer. D.3Gradient-Based vs. Gradient-Free Weight Learning D.3.1Layer Weight Analysis Figure 8 shows the perlayer adapter weights learned by GRPO for a single target author (Robert W. Chambers, 푘=4). The weights are far from uni form: adapter PGW (P. G. Wodehouse) receives strong positive weights throughout, while adapters VW (Virginia Woolf) and GO (George Orwell) are actively suppressed with negative weights (down to −1.0) at deep layers 24–31 — the layers that tend to encode discourselevel and syntactic pat terns (Jawahar et al., 2019). The early layers (0– 10) show moderate mixedsign weights, and the middle layers (14–21) are dominated by strong positive weights for PGW and VW. This is prin cipled: following the taskarithmetic framework (Ilharco et al., 2023) and TIESMerging (Yadav et al., 2023), LoRA adapters act as task vectors in parameter space, and subtracting a task vector steers the model away from the style that adapter encodes. Figure 9 breaks down the GRPO layerwise weights by 푘 (number of adapters). The pattern de scribed in the main text — positive middle layers, negative deep layers — is consistent across all 푘 0.00.20.40.60.81.0 MIS Score 0 2 4 6 8 Density 0.30.50.60.70.80.9 AuthorMix (mu=0.833, sigma=0.059, n=100) ASTRAPOP-SFT (mu=0.634, sigma=0.115, n=100) ASTRAPOP-DPO (mu=0.631, sigma=0.107, n=100) TinyStyler (mu=0.752, sigma=0.071, n=100) TinyStyler-Sim (mu=0.803, sigma=0.072, n=100) STYLL (mu=0.679, sigma=0.099, n=100) FewShot (mu=0.814, sigma=0.085, n=100) Neutral (mu=0.786, sigma=0.099, n=100) Figure 6: Kernel density estimate (KDE) of MIS score distributions for all methods. Legend shows the mean (휇), standard deviation (휎), and number of evaluation pairs (푛=100) for each method. Vertical lines mark the mean of each distribution. Figure 7: LoRAHub layerwise: global average of learned layer weights across all 10 target authors and all 푘 values. All weights are near zero, indicating that gradientfree optimization does not produce structured layerwise patterns. Vertical lines indicate ±1 standard deviation. values (푘=2–10), confirming that this structure is a robust property of GRPO optimization rather than an artifact of a specific 푘. In contrast, Figure 7 shows the LoRAHub (gra dientfree) layerwise weights averaged across all 푘 values and target authors. The weights are near zero across all layers, with no discernible struc ture — confirming that gradientfree optimization struggles to exploit the layerwise search space effectively. D.4Analyzing Resource Consumption Wall-Clock Training Cost: To facilitate fair effi ciency comparisons, we report wallclock training cost per target author — the total compute re quired before the system can serve a single new target individual. This is the practically relevant Figure 8: Perlayer adapter weights learned by GRPO for target author Robert W. Chambers (푘=4). Each col umn is one of the top푘=4 stylesimilar related authors (PGW, VW, OW, GO); each row is a transformer layer (0–31). Red = positive weight (adapter is drawn upon); blue = negative (adapter is suppressed). unit: it measures how quickly a user can be on boarded. For AuthorMix, the cost to adapt to one target author has two components: 1. Adapter training (one-time). Each high-re- source author’s LoRA adapter is finetuned independently in ∼2 h on a single A100 GPU. With 푘 related authors, this costs 푘×2 h. These adapters are trained once and reused for any future target that selects them — so this cost is shared across targets. For the first Figure 9: GRPO layerwise: global average of learned layer weights per 푘 (푘=2–10), averaged across all 10 target authors. Each subplot shows one 푘 value. Red = positive weight; blue = negative weight. Vertical lines indicate ± 1 standard deviation across target authors. target using a given adapter set it must be paid in full; for subsequent targets that reuse the same adapters, it is zero. 2. Weight learning (per target). Learning the mixing weights via GRPO takes ∼40 min per target author on a single A100 GPU. This is the only cost that recurs for each new target. 3.Inference cost remains unchanged across all AuthorMix variants, regardless of 푘, since the mixed adapters are merged into the base model; it is therefore comparable to that of baselines built on the same LLaMA3.18B Instruct backbone. The total cost for the first target at 푘=4 is there fore 4×2+0.67≈8.7 h. Each additional target that reuses the same adapter library requires only ∼40 min. For the baselines, the entire training cost must be invested before any target can be served: • ASTRAPOP-SFT: ∼10 h of full SFT fine tuning (onetime). • ASTRAPOP-DPO: ∼48 h of SFT + DPO training (onetime). • TinyStyler: reconstruction pretraining on ∼8M texts, full SFT, distillation. Despite using a small 800Mparameter backbone, TinyStyler’s training cost remains high (>48 GPU h) because it requires fullmodel fine tuning (reconstruction plus selfdistillation) and is likely higher than both ASTRAPOP and our method. • STYLL / Few-shot: No training cost (infer enceonly via API). Once trained, baseline methods adapt to new targets at zero additional cost (by changing the prompt or embedding). AuthorMix instead pays ∼40 min per new target — but requires 5.5× less compute than ASTRAPOPDPO and far less dataset than TinyStyler to reach the first target. D.5Output Examples D.5.1MIS output analysis Table 1 shows outputs of different styletransfer methods for a single sentence along with their mutual implication score (MIS) that captures how well the meaning is preserved. The original con veys three key elements: the speaker (“I”), the intent (“to tell something”), and the addressee (“you all”). Only AuthorMix preserves all three. ASTRAPOP changes the speaker to “he” and drops the group address; STYLL preserves the speaker but shifts “you all” to “everything,” los ing the direct address; TinyStyler preserves the content but degenerates into ungrammatical repe tition. GPT5.1 retains all three elements but adds embellishment not present in the original (“I broke in at last,” “lay it bare”). The neutral baseline depersonalizes the speaker entirely (“the speaker,” “they”) and removes the addressee. D.6Qualitative Examples Table 8 by fixing a single target author (Fitzgerald) and varying the source text, revealing how each method handles different inputs when targeting the same style. Target Style Reference texts (Fitzgerald): “That’s what you ought to do,” she exclaimed triumphantly.“, “I can just see you,” she stormed, “letting him back you down!”, “Sit down and wait,” suggested the lieutenant nonchalantly.... MethodOutput (target: Fitzgerald) Source (Widger): ‘I cannot do it,’ said Ethelberta; ‘and I will not.’ AuthorMix“No, I won’t do it,” Ethelberta declared firmly. “Nor will I, under any circumstances.” ASTRAPOPDPO“I couldn’t do it,” Ethelberta said, “and I don’t want to do it. TinyStylerEthelberta said, “I cannot do it, and I will not do it.” TinyStylerSimEthelberta said, ‘I cannot do it, and I will not.’ STYLLEthelberta declared she couldn’t and wouldn’t even try. Few shot (GPT5.1)“I can’t do it,” Ethelberta declared; “and I won’t.” Source (Widger): ‘Well; better late than never, if so be he’l have her now.‘ AuthorMix“Better late than never, I suppose, and if he’s going to have her, it’s just as well he does it now.” ASTRAPOPDPO“Yes, it’s better for him to have her now, even if it is late. TinyStyler‘Well, better late than never, if he has her now.’ TinyStylerSim‘Well, better late than never, if he’l have her now, if he doesn’t have her now...‘ STYLL(Roguishly) Better late than never, right? Few shot (GPT5.1)“Well, it’s just as well he’s come around at last—provided he’s still willing to take her.” Source (Widger): Darton held his breath till he said with treacherous solicitude, ‘Try again—’tis coyness.‘ AuthorMixHe held his breath, and then whispered with a treacherous mixture of concern and flirtation, ‘Try again, it’s just coyness.‘ ASTRAPOPDPODarton kept quiet until he spoke, as if she had been coy about it and he was urging her to be bold again. TinyStylerDarton was a little squirming in his seat until he said with treacherous solicitude, ‘Try again.’ TinyStylerSimDarton held his breath until he said with a treacherous solicitude, “Try again—‘tis coyness.“ STYLLDarton kept mum until he finally broke the silence, hinting that their hesitance was just a coy act, and egging them on to have another go. Few shot (GPT5.1)Darton, still holding his breath, murmured with a deceitful gentleness, “Try once more—it’s only shyness.” Source (Widger): One of them actually said that you must be fifty to have got such an experience. AuthorMix“If you’ve had an experience like that, you must be at least fifty,” he said bluntly. ASTRAPOPDPO“He said a man’s got to be fifty years old to have that much experience.” TinyStylerOne of them said you had to be fifty to have such an experience, smh. TinyStylerSimOne of them said that you must be 50 to have had such an experience ... STYLLIt was said that you needed to be half a century old to have gathered so much experience. Few shot (GPT5.1)One of them even ventured the opinion that I must be fifty years old to have managed such an adventure. Table 8: Multiple source sentences rewritten into Fitzgerald’s style by all methods. AuthorMix produces stylistically distinct outputs while preserving meaning across different source inputs for the same target author.