InMyStyle: Privacy-First AI Text Rewriting with Small Language Models

An academic paper published on arXiv (2607.29238) presents InMyStyle, a privacy-centric framework that modifies small language models to rephrase AI-edited content, aligning it with a user's unique writing style without needing an instruction prompt during inference. The approach employs several local helper LLMs to generate paired training examples from user documents and fine-tunes LoRA adapters on base models with parameters ranging from 0.5B to 7B. It accommodates varying input lengths through length-aware generation budgets and automatic chunking. Testing on 219 pairs from a scientific-paper dataset revealed that the automatic composite score stabilizes at 0.69 (on a scale of 0-1) across all model sizes with both greedy and sampled decoding. This suggests that smaller models are adequate for the rewriting task, with size influencing trade-offs instead of a consistent quality hierarchy. In a secondary assessment, 400 ratings from five LLM judges indicated that InMyStyle's outputs had an average perceived AI-ness score over 20% lower than the baseline, reflecting a decreased detectability of AI-generated text. The system is tailored for individual users, emphasizing privacy, and operates solely on local models.