Search papers, labs, and topics across Lattice.
This paper introduces UNIFUSION, a method that adapts autoregressive language models to uniform-noise diffusion by establishing a unified reverse-rate objective that connects various existing models. By deriving a shared interface for switching between mask and uniform kernels, the authors enable the continual pre-training of GPT2 checkpoints specifically for uniform-noise diffusion. The results demonstrate that UNIFUSION outperforms existing models in generative perplexity and unigram entropy, achieving superior accuracy on benchmark tasks at both 124M and 355M parameter scales.
UNIFUSION achieves unprecedented performance in generative tasks by seamlessly adapting autoregressive models to uniform-noise diffusion, outperforming all evaluated models on key metrics.
Existing methods mainly adapt pretrained autoregressive (AR) language models to masked diffusion, whereas we directly adapt them to uniform-noise diffusion, where every token remains editable during sampling. However, adapting AR checkpoints across corruption kernels remains challenging because existing DLMs use different objectives and prediction parameterizations. We establish connections among SEDD, MDLM/GIDD, M2S, and Neural CTMC by expressing their conditional losses as a single generalized Kullback--Leibler objective over model reverse rates. We further derive conversions from clean-token predictions to concrete-score, posterior-mean, and exit-rate/jump parameterizations, yielding a shared \(x_0\) interface that supports switching between mask and uniform kernels. Building on these connections, we propose \ours{}, a simple continual pre-training approach for directly adapting pretrained GPT2 checkpoints to uniform-noise diffusion. Through systematic evaluation of 124M- and 355M-parameter models, we show that \ours{} steadily improves the trade-off between generative perplexity (GenPPL) and unigram entropy as the sampling budget increases from 16 to 256 steps. At 256 steps, \ours{}-S and \ours{}-M achieve GenPPL/entropy pairs of \(97.783/5.2626\) and \(71.516/5.6669\), respectively; no evaluated model at the same scale simultaneously outperforms \ours{} on both metrics. At both scales, \ours{} also achieves the highest WinoGrande, SIQA, and BBH accuracy among the compared diffusion models.
Tsinghua AI