The authors investigate fairness in diffusion language models (DLMs), which use iterative denoising for generation instead of autoregressive next-token prediction.
Diffusion model fairness shifts with sampling settings, not just weights.
Saito et al. show that bias magnitude and model rankings vary substantially across block length and remasking strategies, and that debiasing does not stabilize these measurements.
Big Tech
Koshiro Saito · Diasuke Oba · Danushka Bollegala · Naoaki Okazaki
Institute of Science Tokyo · University of Liverpool · Amazon · AIST · NII LLMC
Research Digest··2 min read
Saito et al.
§