Research Summary: Small Is Enough: Per-User Style Rewriting of AI-Edited Text via LoRA Adapters

A recent research development introduces InMyStyle, a privacy-centric, single-user system designed to adapt small language models (LLMs) to rewrite AI-edited text into an individual's unique writing style. This system operates without explicit inference prompts and leverages local helper LLMs to generate training data, subsequently fine-tuning LoRA adapters on Qwen2.5 models. It supports varied input lengths through automatic chunking and budget management, demonstrating effectiveness in a case study on scientific writing.

Why it matters

This research is strategically important as it addresses the growing need for personalized and privacy-preserving AI applications, particularly in content generation and refinement. The ability to customize AI outputs to individual writing styles locally can enhance user adoption and trust in AI tools, while also mitigating data privacy concerns associated with centralized large language models.

Key insights

• InMyStyle is a privacy-first system for personalized AI-edited text style rewriting.
• It uses local helper LLMs to generate paired training examples for fine-tuning.
• The system employs LoRA adapters on small Qwen2.5 models (0.5B to 7B parameters) for style adaptation.
• It operates without requiring an instruction prompt during inference, focusing on implicit style learning.
• Automatic chunking and length-aware generation budgets facilitate processing diverse input sizes.
• A single-user case study demonstrated plateauing composite scores across various model sizes for scientific writing style adaptation.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.29238