Latent diffusion model provides efficient and high-quality paraphrase

Paraphrase generation requires diverse generation of high-quality utterance by the given semantics, which is a challenge for traditional end-2-end text generation.

Inspired by the diffusion modeling for diverse image generations, a research team from Nanjing University led by Wei Zou managed to reconcile the quality and diversity for paraphrase generation via latent diffusion modeling (Latent Diffusion Paraphraser, LDP), and published their new research on 15 January 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed a latent diffusion modeling at encoded text space, which offers a controllable semantic intervention. The off-the-shelf pre-trained text encoder and decoder bridges the diffusion semantic space with the valid text, preventing overheads of diffusion process for straightforward text space. Compared to the straightforward textual diffusion generations, LDP shares the similar efficiency of traditional end-2-end text generation

model.

The diffusion process allows additional semantic controls to ensure the paraphrase quality, where semantic controls are fine-tuned by sampled input segments from training inputs without additional annotations, which is a great convenience.

Experiments verified the LDP in English paraphrase generation on Quora Question Pair, Twitter-URL, PAWS-wiki paraphrase datasets, with a state-of-the-art generation even comparable to the open-source large language model generations, but at much cheaper costs. LDP with input semantic controls also surpass the diffusion baselines in generation quality and efficiency. Further analysis shows that LDP with its controller implementation is also helpful for similar text generations with diversity and quality requirements, such as question generation and domain adaptation.