Differences in Text Generated by Diffusion and Autoregressive Language Models

Empirical comparison shows diffusion language models exhibit lower n-gram entropy, higher semantic coherence, and higher semantic diversity than autoregressive models; controlled experiments attribute coherence gains to DLM training objectives.