Linguistic style, or how something is said, encodes information about authors and communities and shapes how meaning is perceived by humans and by LLMs. Style plays a central role across the full NLP pipeline: the stylistic diversity of training data influences model robustness, generated texts are expected to have a helpful or personalized style, evaluation on texts with limited stylistic variation can hide how a model behaves in the wild, and applications like AI-generated text detection and text anonymization rely on style features. Yet despite rapid advances in language modeling for reasoning and semantics, style remains comparatively understudied. The field lacks a broadly accepted definition and a consensus on how style should be modeled, measured, and applied.
StyleMatters brings together researchers from NLP, computational (socio)linguistics, forensic linguistics, stylometry, and adjacent fields to explore open questions and chart the future of style in NLP. To anchor these discussions in practical progress, the workshop hosts a shared task on style representation learning built on the Style Text Embedding Benchmark (STEB).