Achieves over 10x faster sampling for diffusion language models by shifting the process into continuous semantic space.
March 24, 2026
Original Paper
CRoCoDiL: Continuous and Robust Conditioned Diffusion for Language
arXiv · 2603.20210
AI-generated illustration
The Takeaway
Diffusion language models often suffer from slow, iterative sampling; this unified fine-tuning approach (CRoCoDiL) enables high-quality text synthesis with massive speed gains, making non-autoregressive generation a viable competitor to standard transformers.
From the abstract
Masked Diffusion Models (MDMs) provide an efficient non-causal alternative to autoregressive generation but often struggle with token dependencies and semantic incoherence due to their reliance on discrete marginal distributions. We address these limitations by shifting the diffusion process into a continuous sentence-level semantic space. We propose CRoCoDiL (Continuous and Robust Conditioned Diffusion for Language), a unified fine-tuning approach that jointly trains an encoder-demasker archite