Abstract
Although auto-regressive models excel in natural language processing, theyoften struggle to generate diverse text and provide limited controllability.Non-auto-regressive methods could be an alternative but often producedegenerate outputs and exhibit shortcomings in conditional generation. Toaddress these challenges, we propose Diffusion-EAGS, a novel framework thatintegrates conditional masked language models into diffusion language modelsthrough the theoretical lens of a conditional Markov Random Field. In doing so,we propose entropy-adaptive Gibbs sampling and entropy-based noise schedulingto counterbalance each model's shortcomings. Experimental results show thatDiffusion-EAGS outperforms baselines and achieves the best quality-diversitytradeoff, demonstrating its effectiveness in non-autoregressive textgeneration.