Despite recent progress in LLMs for text style transfer, most existing methods rely on costly task-specific training and offer limited control over separating stylistic modification from content preservation. We propose Diff4TST, a diffusion-based language model that formulates text style transfer as an explicit copy-and-edit process. Built upon masked diffusion language models, Diff4TST introduces a style-aware noise schedule that selectively perturbs stylistic tokens while preserving content-bearing tokens during supervised fine-tuning.At inference time, we further introduce a generate-then-refine strategy that iteratively improves style compliance via gradient-based token re-masking, without reinforcement learning or external reward models. Extensive experiments on both fine-grained and polarity-based benchmarks show that Diff4TST achieves substantially improved style accuracy and controllability while maintaining strong content preservation and fluency. These results suggest diffusion-based language models as a principled and effective alternative to autoregressive pipelines for text style transfer.

Unsupervised Text Style Transfer (UTST) aims to build a system to transfer the stylistic properties of a given text without parallel text pairs.Compared with text transfer between style polarities, UTST for controllable intensity is more challenging due to the subtle differences in stylistic features across different intensity levels.Faced with the challenges posed by the lack of parallel data and the indistinguishability between adjacent intensity levels, we propose a SFT-then-PPO paradigm to fine-tune an LLM.We first fine-tune the LLM with synthesized parallel data.Then, we further train the LLM with PPO, where the rewards are elaborately designed for distinguishing the stylistic intensity in hierarchical levels.Both the global and local stylistic features are considered to formulate the reward functions.The experiments on two UTST benchmarks showcase that both rewards have their advantages and applying them to LLM fine-tuning can effectively improve the performance of an LLM backbone based on various evaluation metrics.Even for adjacent levels of intensity, we can still observe a noticeable stylistic difference among the generated text across these levels.