About Me

Hi, I’m Duc, a PhD student at JGU Mainz (Germany), supervised by Katharina von der Wense. I work on AI safety and NLP: rigorously measuring and mitigating how LLM-based systems encode and amplify social disparities rooted in human diversity, spanning cultural and linguistic variation, across text, vision, and audio. I also study how harms can arise from integrating AI models into real-world pipelines, and how to mitigate them.

• Cultural Variation: Measuring and mitigating gaps in model behavior arising from cultural contexts, conventions, and norms

  • Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision-Language Models @ NAACL 2025
  • On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures @ ACL 2025

• Linguistic Variation: Measuring and mitigating how social meaning in language variation drives discriminatory model behavior

  • Large Language Models Discriminate Against Speakers of German Dialects @ EMNLP 2025
  • Meenz bleibt Meenz, but Large Language Models Do Not Speak Its Dialect @ LREC 2026

• Harm Propagation at Deployment Scale: Measuring and mitigating how integrating models into real-world pipelines creates harm

  • Greater accessibility can amplify discrimination in generative AI (Preprint)
  • From If-Statements to ML Pipelines: Revisiting Bias in Code-Generation @ ACL 2026 Findings

Together, these studies reflect my broader research agenda: building the empirical case that socially-encoded disparities in LLM behavior are a safety risk, one that scales with deployment, and developing targeted interventions that reduce them.

Latest News (See all)

Recent Publications (See all)