Prompt Guided Diffusion for Controllable Text Generation
Mohaddeseh Mirbeygi, Hamid Beigy · 2025
Text generation under control, or producing linguistically coherent and contextually relevant text, has seen tremendous progress thanks to methods based on PPLM, FUDGE, and diffusion-based models.Yet current state-ofthe-art models tend to balance control fidelity with fluency.In addition, classifier-guided strategies (e.g., PPLM) can be predicted in gradient updates providing less coherent text.In contrast, autoregressive-based approaches (e.g., FUDGE) rely on inflexible generation patterns that limit creativity.Recent diffusion methods demonstrate superior performance in iteration and diversity, but indirect methods often fail to introduce sufficient ways to inject taskassociated knowledge, leading to the need for many different complex classifier modules during both training and inference.To address this, we introduce a prompt-guided diffusion framework that seamlessly incorporates structured prompts into the diffusion steps, providing precise and flexible control of the generated text.Each prompt combines a target attribute (for example, a sentiment tag), an example corresponding to that label (for example, a positive review), and a slot for the generated sentence.By encoding such prompts using large pre-trained models (such as BART) and integrating these prompts through cross-attention into the diffusion dynamics, our model achieves new state-of-the-art performance on a variety of tasks ranging from IMDB for sentiment, AG-News for topic, and E2E for structured-output to text.