Diffusion LLM Bias Varies with Sampling Configuration, Study Finds

Researchers at Tokyo Institute of Science show that bias rankings in diffusion language models change with sampling configurations and remain unstable even after debiasing, and they released an interactive demo that scans all sampling combinations.

2026-10-23 ~ 2026-10-23 · 2 related posts

• Diffusion LLM bias rankings shift with sampling configs, persist after debiasing — Bollegala · 2026-10-23
• Interactive demo released for the DLM sampling-fairness paper — Bollegala · 2026-10-23