Token Time Continuous Diffusion for Language Modeling

The paper introduces Token Time Continuous Diffusion (TTCD), a diffusion language model that deterministically maps Gaussian noise to a final token canvas in continuous space, avoiding additional sampling. Its key mechanism assigns each token an individual time, allowing confident tokens to advance faster and enabling differentiated inter-token influence during refinement. A 160M-parameter model trained on OpenWebText and then self-distilled achieves comparable unconditional generation quality at high speedups while outperforming similarly sized discrete diffusion models in conditional generation. The authors report similar gains on Sudoku solving.