# Summary: 2026-07-31_10-14-28Z_SmallIsEnough_Per_UserStyleRewritingofAI_EditedTex.md
Saved: 2026-08-03 02:45
Source: 2026-07-31_10-14-28Z_SmallIsEnough_Per_UserStyleRewritingofAI_EditedTex.md
Model: qwen3.8-flash-next-iq3_xxs

---

## Summary
This paper introduces "InMyStyle," a privacy-first, single-user system designed to adapt small language models (SLMs) to rewrite AI-edited text into an individual user’s unique writing style. The core contribution is the demonstration that fine-tuning LoRA adapters on small models (ranging from 0.5B to 7B parameters) is sufficient for this task, challenging the assumption that larger models are necessary for personalized style transfer. The system operates without instruction prompts at inference time, relying instead on pre-trained adapters derived from the user's own documents. The study highlights a performance plateau in automatic scoring across model sizes, suggesting that model scale primarily affects trade-offs rather than absolute quality in this specific rewriting context.

## Key Contributions
- **Privacy-Centric Personalization:** The system enables per-user style adaptation without sending user data to external servers, addressing privacy concerns inherent in cloud-based AI editing tools.
- **Small Model Sufficiency:** The research demonstrates that small language models (0.5B–7B parameters) achieve comparable performance to larger models for style rewriting, with automatic scores plateauing at 0.69 across all sizes.
- **Reduced AI Detection:** Outputs generated by InMyStyle are perceived as significantly less "AI-like" than the original AI-edited inputs, with perceived AI-ness decreasing as the model size increases within the small model range.

## Methodology
The authors developed a pipeline that uses multiple local helper LLMs to construct paired training examples from a user’s existing documents. These pairs serve as the dataset for fine-tuning LoRA (Low-Rank Adaptation) adapters on base models of varying sizes. The system incorporates length-aware generation budgets and automatic chunking to handle inputs of different lengths. Evaluation was conducted on 219 pairs from a scientific-paper corpus using an automatic composite score. Additionally, a secondary evaluation involved 400 ratings from five LLM judges to assess the perceived "AI-ness" of the outputs compared to the original AI-edited inputs.

## Results
The automatic composite score for style rewriting plateaued at 0.69 (on a 0-1 scale) across all model sizes from 0.5B to 7B parameters, under both greedy and sampled decoding strategies. This plateau indicates that increasing model size does not yield stable quality improvements for this specific task. In the secondary evaluation, InMyStyle outputs received mean perceived AI-ness scores over 20% lower than the helper-AI generated inputs. Furthermore, within the InMyStyle system, larger models within the small model range produced outputs with lower perceived AI-ness, suggesting a nuanced trade-off where model size influences the subtlety of style mimicry rather than overall rewriting accuracy.

## Significance
This work is significant because it challenges the prevailing trend of using large, resource-intensive models for personalized AI tasks. By proving that small models are "enough" for style rewriting, it offers a viable, privacy-preserving, and computationally efficient alternative for individual users. It also provides insights into how model size affects the perception of AI-generated text, which is crucial for applications where human-like authenticity is desired.

## Related Concepts
- LoRA (Low-Rank Adaptation)
- Small Language Models (SLMs)
- Style Transfer
- Privacy-Preserving AI
- AI Detection/Perception
- Fine-tuning
- Local LLMs