Fairness of diffusion language models depends on sampling settings
Beyond Model Weights: Sampling-Dependent Fairness in Diffusion Language Models
Computation and Language
Summary
Diffusion language models are a new way to generate text from computers, different from older models. This study found that the fairness of these models—how much they show social bias—changes depending on how the computers pick words during the generation process. Even when methods to reduce bias are applied, these changes in settings still affect fairness. The authors show that understanding bias in these models requires looking at both the model itself and the way text is generated from it.
What this means in practice
- •For ai system builders: Design and evaluate systems that generate text with awareness of bias sensitivity to sampling configurations in diffusion language models.
- •For ethics compliance teams: Develop more accurate fairness assessments by including sampling-dependent bias effects in language model evaluations.
Authors
Koshiro Saito, Diasuke Oba, Danushka Bollegala, Naoaki Okazaki
Abstract
Diffusion language models (DLMs) are emerging as a practical alternative to autoregressive language models (ARLMs), yet their fairness remains largely uncharacterized. DLMs introduce sampling controls with no direct autoregressive counterpart, such as block length and remasking strategy, that directly set their speed--quality operating point. This raises a basic question: does a DLM's measured social bias remain stable across these settings? Across LLaDA, LLaDA-1.5, and Dream on Open-BBQ, we find that all three models exhibit social bias, but both the magnitude of measured bias and the relative ordering of models by bias are highly sensitive to the sampling configuration. This sampling dependence also changes what successful debiasing means. While mitigation strategies established for ARLMs transfer effectively to DLMs and reduce bias without degrading downstream performance, sampling dependence persists even after mitigation. Our results show that fairness in DLMs cannot be characterized independently of the sampling configuration, even after debiasing. Project page is available at https://pakapaka333.github.io/Beyond-Model-Weights/.