Shows that low-resource authorship style transfer improves when models avoid static author embeddings and instead preserve context-dependent stylistic variation.
HyperStyler is a novel architecture for low-resource authorship style transfer (LAST) that decouples the task into context-aware style selection and style realization to overcome the limitations of static author embeddings. By using a Stylo-navigator to predict dynamic style coordinates based on source context and target references, and a Stylo-hypernet to realize these via parameter modulation, the model preserves fine-grained, context-dependent stylistic variations.
Extensive experiments on Reddit, Blog, and News datasets demonstrate that HyperStyler consistently outperforms existing baselines, including LLM-based approaches and general methods like TinyStyler. It achieves superior style fidelity and semantic preservation while being over 1.8× faster than LLMs at inference and requiring only a 2.4% parameter increase over T5-large, highlighting its efficiency and robust generalization across diverse domains.
HyperStyler addresses low-resource authorship style transfer, where the goal is to rewrite or generate text so that it reflects a target author’s style while preserving the original content, using only a limited amount of target-author data. The material challenges a common assumption in author style modeling: that an author can be represented by a static embedding or fixed style vector. Instead, it frames authorship style as a context-dependent phenomenon and proposes a method for navigating style space in a way that is sensitive to the surrounding linguistic and discursive context.
The central technical contribution is a framework that combines context-aware style navigation with hypernetworks to instantiate style-conditioned transformations dynamically. Rather than collapsing a target author’s stylistic behavior into a single averaged representation, the approach allows the model to adapt its stylistic expression according to the input context. The hypernetwork plays a key role by generating or conditioning model parameters from this context-sensitive style signal, enabling more flexible, instance-specific style transfer without requiring large per-author datasets. The key insight is that authorial style is not a fixed identity; the same author may vary their diction, rhythm, and rhetorical tendencies across topics, registers, and local syntactic contexts, and preserving that variation is important for faithful style transfer.
This matters because authorship style transfer is a practically important but data-constrained problem: real-world author corpora are often small, noisy, or unevenly represented, and static author embeddings can overfit to surface-level averages or produce stylistically flat output. By treating style as a dynamic, context-conditioned process, HyperStyler offers a more robust formulation for low-resource settings. More broadly, the work has implications for controlled text generation and style-aware language modeling, suggesting that effective style control may depend less on learning a fixed “author vector” and more on learning how style should be navigated and instantiated across different contexts.