Masked Diffusion Decoding as $x$-Prediction Flow
This work improves decoding efficiency for masked diffusion language models, which is important for practitioners needing faster text generation without sacrificing quality.
Masked diffusion language models suffer from all-or-nothing token decoding, discarding partial beliefs and causing premature commitments. The authors reinterpret mask prediction as x-prediction to enable a continuous decoding flow, achieving 97% of LLaDA's HumanEval performance with only 25% of the decoding budget.
Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens, but their standard decoder reduces each step to a binary action: a position is either committed to a single token or left fully masked, with no representation of partial belief in between. This all-or-nothing regime discards rich predictive information and forces premature, irrevocable commitments, leading to poor performance under a limited decoding budget. In this paper, we reinterpret mask prediction as clean-state prediction ($x$-prediction) and show that it can be used to induce a continuous flow in input embedding space. Building on this view, we propose a continuous decoding framework for MDLMs where tokens can accumulate partial progress at each diffusion step and remain revisable. To match the uneven contextual constraints across positions in language, we replace the globally synchronous schedule in image diffusion with a confidence-based asynchronous update in which the diffusion progress is token-wise accumulated. Additionally, we introduce a lightweight policy network and formulate its training as a reinforcement learning problem. Applied to pretrained LLaDA, our continuous decoder reaches 97% of its performance on the HumanEval dataset with 25% of decoding budget.