Bias in Large Models
I study stereotypical bias in both large language models and vision-language models, with an emphasis on how these biases surface in downstream tasks.
Research
Research Interests
My research explores trustworthy and context-aware AI systems, with a particular focus on bias in large models, multimodal learning, and practical evaluation for machine learning in real-world settings.
Trustworthy AI
Bias & Fairness
Multimodal Learning
Directions
Current Research Directions
Debiasing Strategies
Beyond measurement, I am interested in methods that can actively mitigate harmful group disparities while preserving useful inference behavior.
Multimodal Context
I explore how visual inputs can be combined with local or environmental context to create more grounded and effective multimodal systems.
Applied Evaluation
I care about evaluation protocols that reflect real use: noisy signals, incomplete context, diverse users, and deployment-relevant decision constraints.
Questions
What I Am Exploring
- How local context changes bias measurements in large language models.
- How vision-language models can be debiased without overcorrecting useful signals.
- How metadata can improve multimodal representation learning in applied domains.
- How evaluation can better reflect fairness and robustness in open-world settings.
- How research outputs can translate into practical, inspectable AI systems.
Methods
Methods & Approach
Machine Learning
Vision-Language Models
Bias Measurement
Dataset Design
Empirical Evaluation
Applied AI Systems
My recent outputs suggest a research style that combines methodological questions with application-aware evaluation: debiasing, local-context analysis, and multimodal modeling under realistic constraints.
Outputs
Selected Research Threads
CABIN
Debiasing vision-language models using backdoor adjustments, published at IJCAI 2025.
Investigating Stereotypical Bias
Master's thesis on bias in large language and vision-language models, completed at the University of Auckland.
LIBRA
A preprint that measures bias in large language models from a local-context perspective.