Diffusion language models (dLLMs) generate text by starting from a fully masked canvas and progressively revealing tokens. A new preprint on arXiv shows that the sampler's role in this process is more consequential than previously thought: it controls not just which token gets written, but when that token is committed. The authors argue that this second decision is a distinct source of bias.
According to the abstract, the timing of token commitment "changes who appears in generated text." This points to a representational skew—some identities or entities may be systematically more likely to be revealed early or late, affecting the final composition of the output. The paper's title, "Who Gets Unmasked?," frames this as a question of fairness and visibility in model outputs.
The finding adds a new dimension to discussions of bias in language models, which have largely focused on token selection probabilities. Here, the order of unmasking itself becomes a bias mechanism. The abstract is brief and does not specify which groups are affected or how the effect was measured, but the implication is clear: even if a model's vocabulary and probabilities are neutral, the sampler's scheduling can still produce biased text.