HyperStyler: Low-resource Authorship Style Transfer via Context-aware Style Navigation and Hypernetworks

• People express the same content through distinctive lexical, syntactic, and expressive choices; LAST rewrites text in an arbitrary author’s style from a few examples while preserving meaning.Unlike traditional style transfer requiring massive corpora, LAST extends the task to everyday writers with only a small number of sentences.Source: §1 · p. 1
• In-context or inference-time control (Patel et al., 2024; Khan et al., 2024; Horvitz et al., 2024a) and pseudo-parallel alignment (Liu et al., 2024; Horvitz et al., 2024b) struggle to jointly preserve style fidelity and semantics.Source: §1 · p. 1
• Author style shifts with topic and register, so unprioritized references can dilute style into an average or overemphasize a salient example.The paper identifies both compressing references into one static embedding and feeding all references without prioritization as sources of this problem.Source: §1 · p. 1
• Hidden-state style injection entangles stylistic signals with semantic content, hampering precise realization of unseen authors’ diverse styles while preserving meaning.This interference is especially problematic for LAST because target authors have open-ended stylistic variation rather than fixed style categories.Source: §1 · pp. 1–2
• Across Reddit, Blog, and News, HyperStyler outperforms baselines including LLM-based approaches, generalizes across domains, runs over 1.8× faster than LLMs, and performs better with 2.4% more parameters than T5-large.The paper presents these findings as evidence of practical utility for time- and resource-constrained applications.Source: §1 · p. 2
• HyperStyler separates style selection from realization: Stylo-navigator predicts style coordinates from source context and target-author references, and Stylo-hypernet realizes them through dynamic parameter modulation.The parameter-space control is intended to reduce content-style entanglement and enable fine-grained controllability.Source: §1 · p. 2

2.1 Unsupervised Alignment for LAST

• HyperStyler follows two-stage unsupervised alignment but aligns toward predicted style coordinates rather than a fixed author embedding, accounting for stylistic variation.Prior methods reconstruct text from style-neutralized paraphrases conditioned on static author embeddings or reference samples, then align using filtered pseudo-parallel data (Krishna et al., 2020; Horvitz et al., 2024b; Liu et al., 2024).Source: §2.1 · p. 2

2.2 Hypernetworks

• HyperStyler addresses underexplored few-shot control of fine-grained linguistic patterns for unseen authors by conditioning hypernetworks on reference-derived style coordinates.Hypernetworks generate target-model parameters from external signals and have mainly supported task- or domain-conditioned adaptation; this approach enables dynamic style control while reducing content-style entanglement.Source: §2.2 · p. 2

3.1 Overview

• HyperStyler separates meaning-preserving author-style rewriting into context-and-reference-based style selection and stylistic realization, using Stylo-navigator and Stylo-hypernet modules on an encoder–decoder paraphraser.Source: §3.1 · p. 2

3.2 Stylo-navigator

• The Stylo-navigator embeds references in a content-independent STYLE space, combines reference self-attention with source-conditioned cross-attention, and interpolates context-dependent weights into a style coordinate.Source: §3.2 · p. 3

3.3 Stylo-hypernet

• Stylo-hypernet uses style-conditioned, layer-specific signals to generate decoder cross-attention key/value prefixes and low-rank FFN updates through target-specific MLPs.Source: §3.3 · pp. 3–4

3.4 Training Procedure

• Training first builds a semantically reliable paraphrasing backbone from synthetic paraphrase pairs using bidirectional reconstruction, then freezes it while jointly training navigator and hypernet on unsupervised reconstruction.Source: §3.4 · p. 4
• In Stage 2, the navigator learns to identify the source-matching reference, while teacher-forced ground-truth style embeddings train the hypernet reconstruction without propagating navigator errors.Source: §3.4 · p. 4
• Stage 3 aligns for transfer by self-distilling and filtering pseudo-parallel outputs with predicted z for style-fidelity evaluation, then jointly training navigator and hypernet.Source: §3.4 · p. 4

4.2 Results

• HyperStyler achieves the highest JOINT scores across Reddit, Blog, and News by improving TOWARDS while maintaining competitive SIM; reranking brings further improvement.On News, where most baselines perform relatively poorly, HyperStyler also outperforms all baselines, including LLMs.Source: Table 1 · p. 5, §4.2 · p. 6
• TinyStyler degrades when transferring from News to Reddit and Blog, where target-author inter-author distances are approximately 2.04× and 1.68× larger than in-domain, respectively; HyperStyler is consistently strong across most domain pairs.Source: §4.2 · p. 6

4.3 Analysis and Ablation Study

• Predicted z reaches 0.82 cosine similarity and 0.80 MRR versus mean pooling’s 0.58 and 0.21; explicit source-conditioned selection beats mean pooling and implicit cross-attention on TOWARDS.HyperStyler also remains robust as target-author style variation increases, unlike degrading baselines.Source: §4.3 · p. 7, Figure 2 · p. 7, Figure 3 · p. 7
• HyperStyler’s TOWARDS/SIM ratio is 0.263 versus 0.192 for layer-wise hidden-state injection, while global injection fails to induce style changes.The comparison indicates parameter modulation achieves higher style fidelity for the same semantic cost than hidden-state injection.Source: §4.3 · pp. 7–8
• Jointly modulating FFN and cross-attention best balances style fidelity and content preservation; modulating self-attention breaks sentence structure.Across rank and prefix-length combinations, both modulation components outperform variants removing either one; increasing rank offers limited benefit and longer prefixes do not guarantee gains.Source: §4.3 · p. 8, Figure 4 · p. 8

4.4 Human Evaluation

• Human evaluation finds HyperStyler has the highest SF and G.mean; its SF significantly exceeds GPT5.4 and ParaGuide, while its CS is not significantly different from ParaGuide.ParaGuide has the highest CS score.Source: §4.4 · p. 8, Table 5 · p. 8

4.5 Efficiency

• With rank and prefix length set to 1, HyperStyler still outperforms TinyStyler while adding only 2.4% of the underlying model’s parameters.Reducing rank and prefix length slightly degrades performance.Source: §4.5 · p. 8, Table 6 · p. 8
• HyperStyler runs over 1.8× faster than open-source LLMs and over 2.0× faster than API-based LLMs, using less than one-eighth of their VRAM.It takes approximately one second per inference on a single A100 GPU and remains faster than LLM baselines with reranking.Source: §4.5 · p. 9

• HyperStyler decouples style selection from stylistic realization to address mode averaging and style-content entanglement.It selects the most contextually appropriate style from limited references and realizes it through parameter-space modulation.Source: §5 · p. 9
• HyperStyler consistently outperforms existing baselines and generalizes robustly across diverse domains.Source: §5 · p. 9
• 2.4% increase in parameters still allows HyperStyler to maintain superior performance.Source: §5 · p. 9

Limitations

• The study follows short-text LAST protocols, typically transforming one to three sentences, while long-form transfer requires modeling discourse structure, argument development, coherence, transitions, and narrative flow.Current UAR-based metrics are primarily validated on short texts and may miss cross-sentence stylistic coherence, motivating long-form evaluation protocols.Source: Limitations · p. 9
• Experiments use English corpora; multilingual and cross-lingual transfer requires language-appropriate style representations and reliable evaluation of style fidelity and content preservation.HyperStyler is not inherently language-specific, but these extensions remain future work.Source: Limitations · p. 9

Ethics Considerations

• HyperStyler’s high-fidelity stylistic imitation could enable unauthorized impersonation and challenge existing AI-text detection, so the authors call for respectful use and user responsibility.They argue sophisticated synthetic text requires a new paradigm for authorship-aware detection.Source: Ethics Considerations · p. 9
• Because Reddit training data was not explicitly pre-filtered, the model may generate unethical or biased outputs; the authors recommend safety filters and post-processing for deployment.The data may contain offensive language, sexual content, or unethical sentiments.Source: Ethics Considerations · p. 9

A.1 Reddit

• The Million User Dataset (Khan et al., 2021) is a publicly available Apache-2.0 Reddit corpus of over 300 million posts by approximately one million users across one year.Evaluation uses three predefined splits from Patel et al. (2024).Source: §A.1 · p. 13
• Reddit evaluation includes Diverse authors across at least 13 subreddits, Random authors, and Single authors whose posts all come from a college-football subreddit.The splits are predefined in Patel et al. (2024).Source: §A.1 · p. 13

A.2 Blog

• The Blog Authorship Corpus contains posts by 19,320 bloggers and is freely available for non-commercial research.It was collected from blogger.com.Source: §A.2 · p. 13

A.3 News

• The All-the-news dataset comprises articles collected from major U.S. and English-language news outlets.Source: §A.3 · p. 13
• For single-author analysis, the News data excludes articles with missing author information, organizational or other non-individual authors, and multiple authors.Only articles attributed to clearly identifiable individuals are retained.Source: §A.3 · p. 13
• Training dataset statistics are presented in Table 7.Source: Table 7 · p. 13

A.4 Stylistic Distance across Datasets

• Table 8 measures intra-author style variation as mean cosine distance to author centroids and inter-author distance as mean pairwise cosine distance in UAR space.Higher values indicate greater within-author stylistic variability and greater separation between authors.Source: §A.4 · p. 13, Table 8 · p. 13

A.5 Cross-domain Inter-author Distance

• Cross-domain inter-author distances exceed in-domain distances across all domain pairs, with the largest relative differences for News-to-Reddit and News-to-Blog.Because TOWARDS is normalized by source-target distance, TOWARDS and JOINT should be compared across models within the same domain pair, not across pairs with different distances.Source: §A.5 · p. 13, Figure 5 · p. 13

B Implementation Details

• HyperStyler uses T5-large with pre-norm, 16-head attention, d = de, and projection matrices shared across embedding tables.Training uses two NVIDIA A100 80GB GPUs, batch size 128, lowest-validation-loss checkpoints, and inference sampling at temperature 0.8 with top-p 1.0.Source: §B.1 · p. 14, B Implementation Details · p. 14

B.1.1 Selection of Style Embedding Space

• The STYLE embedder provides content-independent style representations and enables comparison with TinyStyler and ParaGuide, which also use it for style conditioning.It is contrastively trained with negative samples from the same topic and domain; Reddit clustering showed interpretable stylistic patterns beyond content.Source: §B.1.1 · p. 14

B.1.2 Selection of Activation Function for FFN Adapter

• Activation-function adapters generally outperform the no-activation variant overall, whose strong style transfer is offset by weaker semantic preservation.Differences among activation functions are small; HyperStyler uses GeGLU because it is competitive and is also used in the underlying T5-v1.1large FFN.Source: §B.1.2 · p. 14

B.2 TinyStyler

• TinyStyler uses its original configuration and training code, with its public Reddit checkpoint substituted because local retraining performed worse than originally reported; StyleMC was reimplemented from its paper and adapted to the STYLE embedder.For StyleMC, author embeddings were mean-pooled to align with the evaluation protocol used for other models.Source: §B.2 · p. 14, §B.3 · p. 15

B.4 ASTRAPOP

• Despite a more than 9× model-size difference, HyperStyler achieves higher JOINT scores than ASTRAPOP with its original LLaMA-2-7B backbone across three datasets.No ASTRAPOP configuration performs consistently best across datasets, so the paper reports the two variants preserving the original input order in Table 1.Source: §B.4 · p. 15

B.5 Paraguide

• ParaGuide follows the experimental setup of its original paper.The appendix states that the setup follows Horvitz et al. (2024a).Source: §B.5 · p. 15

B.6 In-context Learning Methods

• GPT-based models and Llama-3.1 use Horvitz et al.’s (2024b) prompt with default API settings for GPT-4-turbo, GPT-5-mini, and GPT-5.4.STYLL follows Patel et al.’s (2024) original setup using Qwen2.5-7B.Source: §B.6 · p. 15
• Paraguide’s hyperparameters are presented in Table 14.The supplied passage identifies the table but does not provide its hyperparameter values.Source: Table 14 · p. 16

C.1 Metric Formula Definition

• For LAST, Patel et al. (2024) define author post sets, transferred-post sets, UAR embeddings, and a similarity function S scaled to [0, 1].The complement is Sc(u, v) = 1 − S(u, v).Source: §C.1 · p. 16
• AWAY measures how far style-transferred text departs from the source author’s style.Source: §C.1 · p. 16
• TOWARDS measures how far style-transferred text moves toward the target author’s style.Source: §C.1 · p. 16
• SIM measures preservation of source-text meaning, using the average Mutual Implication Score between post sets; MIS is attributed to Babakov et al. (2022).Source: §C.1 · p. 16

C.2 Details on Human Evaluation

• Annotators from English-speaking countries with at least 95% approval had to correctly identify the author in all three qualification items to participate.Each item showed five reference texts from each of two randomly selected authors and a held-out target text.Source: §C.2 · p. 16
• For each baseline category, the best-performing model was evaluated on 100 Reddit test-set source-target author pairs, with three annotators per output.Source: §C.2 · p. 16
• Examples containing violent, sexually explicit, or profane content were excluded, and annotators were warned about potentially harmful content before consenting to participate.Source: §C.2 · p. 16
• Annotators were paid 60 cents per pair, based on estimated completion time and an average hourly rate.Source: §C.2 · p. 16
• For style fidelity, annotators compared a transferred text with an anonymized source in randomized order against eight target-author references; majority vote set the label (Krippendorff’s α = 0.10).A score of 1 was assigned when the transferred text was selected, and 0 otherwise.Source: §C.2 · p. 16
• For content similarity, annotators rated source-transfer semantic similarity from 0 (Not Similar) to 2 (Similar), with the mean rating used as the final score.The question and scale were adapted from Liu et al. (2024); instructions appear in Table 16.Source: §C.2 · p. 16, Table 16 · p. 17
• McNemar’s test assessed style-fidelity differences and paired t-tests assessed content-similarity differences, using p < 0.05 as the significance threshold.Style-fidelity instructions appear in Table 15.Source: §C.2 · p. 16, Table 15 · p. 17

D.1 Analysis on the Number of References

• From K = 6, HyperStyler consistently outperforms TinyStyler, and K = 9 references surpasses TinyStyler’s best performance at K = 16.Figure 6 shows JOINT score trends on the Reddit Random split, with standard-error bands over five random samplings.Source: §D.1 · p. 17, Figure 6 · p. 17
• The results suggest that selecting stylistic evidence relevant to source context matters more than simply increasing the number of references.The analysis attributes HyperStyler’s more effective use of references to context-aware style selection.Source: §D.1 · p. 17

D.2 Computational Cost Analysis

• Inference time averages 300 Reddit test instances, while VRAM measures peak FP32 memory for locally hosted models on one NVIDIA A100 GPU.For API-based LLMs, only wall-clock latency is reported; reranking time includes candidate generation and reranking.Source: §D.2 · p. 17, Table 17 · p. 17
• Training-time figures characterize practical costs under each method’s setup, not a controlled efficiency comparison, and omit in-context learning baselines.Methods differ in training objectives, optimization hyperparameters, and schedules.Source: §D.2 · p. 18, Table 18 · p. 18
• HyperStyler’s three stages take 28h, 16h, and 0.5h, respectively, on the Reddit dataset.These reported training times reflect the study’s experimental configuration.Source: §D.2 · p. 18, Table 18 · p. 18
• HyperStyler’s final self-distillation stage uses roughly 40K filtered instances, similar to TinyStyler’s, but takes less training time under the reported configuration.The comparison is specific to the methods’ configurations and is not a strictly controlled training-efficiency comparison.Source: §D.2 · p. 18

D.3 Qualitative Analysis

• Table 20 illustrates target-specific cues—including questions, ellipses, warmth, and concise or punctuation-heavy phrasing—while preserving each source context.The examples vary the output for each source according to the target-author references.Source: §D.3 · p. 18, Table 20 · p. 20
• When source information density conflicts with target style, HyperStyler may omit secondary details under compressive styles or add unsupported content under elaborative styles.The first failure occurs with long, information-dense sources and short, reaction-oriented references; the second with short, self-contained sources and expressive, elaborative references.Source: §D.3 · pp. 18–19, Table 19 · p. 18

D.4 Analysis on Style-dependent and Layer-wise Modulation

• Stylo-hypernet uses bilinear interactions between style coordinates and learnable layer embeddings to produce compatibility scores for decoder-layer modulation targets.The scores are analyzed across style clusters and decoder layers.Source: §D.4 · p. 19

E Licenses and Use of Artifacts

• The paper reports using artifacts consistently with their licenses and intended uses, restricting the CC BY-NC-SA 4.0 artifact to non-commercial research and derived materials to research use.It also states that artifacts are not redistributed contrary to their original licenses or access conditions.Source: E Licenses and Use of Artifacts · p. 19, Table 27 · p. 23

D.4 Analysis on Style-dependent and Layer-wise Modulation

• Figure 7 shows compatibility scores differing across style clusters within a layer and across layers within a cluster, confirming style- and layer-dependent modulation.Scores were averaged within clusters and normalized per layer by the maximum absolute value across clusters.Source: Figure 7 · p. 19, §D.4 · p. 19

D.3 Qualitative Analysis

• Figure 8 shows HyperStyler outputs aligning with distinct target-author styles, while static-embedding and reference-input baselines show weaker separation or alignment.The comparison uses t-SNE projections of outputs for the same source texts transferred to two target authors.Source: §D.3 · p. 18, Figure 8 · p. 20