Title: Short-form Text Rewriting with Phi Silica

Abstract:

Short-form text rewriting represents a restricted form of paraphrasing, characterized by high semantic density and constrained context, which significantly limits opportunities for variation. Although large language models excel at general paraphrasing, small language models (SLMs) frequently face challenges regarding semantic accuracy and resistance to hallucinations when applied to short-form content. This study empirically investigates the adaptation of the SLM, Phi Silica, for short-form rewriting tasks. Our methodology encompasses dataset curation, prompt distillation, parameter-efficient fine-tuning, and comprehensive evaluation. We assembled a corpus of concise, presentation-style text derived from public slide decks, leveraging GPT-5-chat to produce rewrite supervision and to serve as an LLM-as-a-judge for assessment. The experimental outcomes indicate that fine-tuning enhances semantic fidelity, mitigates hallucinations, and yields a higher preference win rate compared to rewrites generated by GPT-5-chat. These results imply that specialized adaptation strategies for SLMs can significantly reduce the performance disparity with cloud-based models, offering actionable insights for deploying SLMs in rewrite tasks that demand high precision.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC