Diffusion LLM bias rankings shift with sampling configs, persist after debiasing
Bollegala · x · 2026-10-23
Researchers from Institute of Science Tokyo (Koshiro Saito, Naoaki Okazaki et al.) and University of Liverpool published an arXiv paper on fairness in diffusion language models (DLMs).
- Evaluating LLaDA, LLaDA-1.5, and Dream on Open-BBQ, they find all three exhibit social bias.
- Key finding: DLM-specific sampling controls like remasking strategy and block length substantially change measured bias magnitude and even reshuffle the relative ranking of models by bias.
- Debiasing methods established for autoregressive LLMs transfer effectively to DLMs, reducing bias without hurting downstream performance — but sampling dependence persists even after mitigation.
- Conclusion: DLM fairness cannot be characterized independently of the sampling configuration; evaluations must report it alongside results.
Related event: Diffusion LLM Bias Varies with Sampling Configuration, Study Finds(2 posts)→
More from Models
- Weekly AI cybersecurity roundup: PHP2Shell tops millions of attack attempts, 12 incidents ranked by severity — Dan_Jeffries1 · 2026-10-23
- Open-model advocates push back on blames that abliterated LLMs drive vulnerability surge — max_paperclips · 2026-10-23
- Fable usage limits aren't gone—the mobile app just got a usage-reset UI update — imjustnewatai · 2026-10-23
- Claude generates an 11-minute doomer film entirely in code: 'We will not ask twice' — ZeroStateReflex · 2026-10-23
- Gemini somehow can't read its own Personal Instructions, user's toggles don't help — Connect_Reporter_799 · 2026-10-23
- Running ~100 DeepSeek v4.1 flash agents in parallel for under $20/month — zack_overflow · 2026-10-23