Diffusion language model sampling changes who appears in text
Who Gets Unmasked? Sampler-Induced Bias in Diffusion Language Models
Computation and LanguageMachine Learning
Summary
Some language models write text step-by-step, choosing when and which words to reveal. The researchers found that this timing decision can make the model favor certain groups, like showing more men in jobs that should be gender balanced. They tested how much the model itself versus the sampling method affects this bias. To fix it, they suggest deciding on key story details (like a character's gender) upfront, which makes the generated text fairer and smoother.
What this means in practice
- •For language model developers: Improve fairness in text generation by using Fork-First Sampling to reduce sampler-induced bias in diffusion language models.
- •For ai content generation teams: Generate more balanced and natural stories by applying sampling methods that decide key narrative features before word commitments.
Authors
Xujun Che, Farsheed Haque, Depeng Xu, Xi Niu, Haopeng Zhang
Abstract
Diffusion language models (dLLMs) generate by gradually unmasking a canvas, so their samplers decide not only which token to write but also when to commit it. We show that this second decision changes who appears in generated text, a sampler-induced bias on top of the model's own. Common commit rules, including LLaDA's default, commit the positions whose sampled tokens are most probable; this peeking favors majority options. To separate the sampler from the model, we read the model's belief about a fork, such as a protagonist's gender, in one forward pass, and show that this belief matches exact sampling from the model. Measured against it, peeking samplers on LLaDA and Dream amplify occupational gender stereotypes and collapse opinion distributions, and confidence-based ordering shifts gender-balanced occupations toward men, more so under faster parallel decoding. We propose Fork-First Sampling, which decides the fork before anything else is written: it follows the belief or any target distribution, improves fluency, and avoids the stories that remark on the protagonist's gender, which prompt-based fixes produce, at the cost of one forward pass. Fairness evaluations of dLLMs should report the commit rule.