A new tool fine-tunes small AI models on a user's own writing so they can rewrite AI-generated drafts to sound like the user wrote them, and it turns out the small models work just fine.
The system, called InMyStyle, is built to run privately: it uses local helper LLMs to generate training examples from a user's own documents, then fine-tunes LoRA adapters on Qwen2.5 models ranging from 0.5 billion to 7 billion parameters. In a single-user case study, researchers tested the adapters on 219 evaluation pairs drawn from 73 paragraphs of one author's scientific writing, training every model size with the same rank-8, three-epoch recipe. A composite quality score, measured on a 0 to 1 scale, landed between 0.689 and 0.695 across all model sizes, with confidence intervals overlapping enough that bigger models showed no real advantage.
That is a useful, if modest, finding: for this kind of rewriting, you don't need a large model chewing through your documents, which matters if the whole pitch is keeping your writing off someone else's servers. But the gains lean toward making text sound less like an AI wrote it, not toward genuinely capturing how the author writes: an authorship classifier's confidence stayed parked near its 0.51 to 0.53 decision boundary, and stylometric improvement was close to zero.
In a secondary test, five LLM judges rated outputs as sounding more than 20 percent less AI-generated than the drafts they started from, with perceived AI-ness dropping further on larger models. That is a solid trick for smoothing over AI-edited prose. Whether it amounts to personalization, or just a well-mannered AI detector evasion tool with better branding, is a different question, and one this single-author study cannot answer.