3 questions for...
Dr. Yao Rong
Groupleader
Dr. Yao Rong is an expert in Explainable AI (XAI), specializing in the development of actionable explanations that move beyond mere model interpretation. Her research focuses on operationalizing model explanations to help users audit, collaborate with, and refine large foundation models in practical use cases such as medical applications. By leveraging explanations as a diagnostic tool, her work explores how to guide large model training and optimize inference performance. Furthermore, her research integrates principles from human psychology to ensure that complex model behaviors remain transparent, reliable, and aligned with human intent.
Dr. Rong earned her Ph.D. in Computer Science from the Technical University of Munich in 2024. Prior to her faculty appointment, she was as a Junior Research Fellow and Future Faculty Fellow at Rice University. Her work is widely published in premier international machine learning venues and addresses the critical need for user-centric AI evaluation. In recognition of her research contributions, she was selected as an MIT EECS Rising Star in 2025.
1.
Trust, transparency and fairness are often mentioned together - but in practice, they can also conflict with each other. How can these goals be balanced?
Exposing counter-intuitive model logic can erode user trust, while high transparency may risk users manipulating the system to achieve specific outcomes, which compromises societal fairness. To balance these, I think domain experts should audit explanations to ensure they are robust and ethically aligned. This approach treats model explanation as a diagnostic tool for model improvement, using an expert-in-the-loop to ensure systems remain transparent and trustworthy.
2.
How can explainable AI methods help to identify errors, biases or uncertainties in models at an early stage?
Model-centric XAI algorithms reveal the features that drive decisions. By visualizing feature importance and causal relationships, we can detect spurious correlations, where a model relies on irrelevant artifacts. Data-centric explanations identify training samples that may be biased or insufficient. Together, these methods reveal underlying problems and facilitate expert feedback to improve model training at an early stage.
3.
Your research bridges the gap between fundamental research and real-world application scenarios. In your view, what is crucial for research results to find their way into practice?
In my view, the most critical factor for translating research into practice is identifying the functional gaps in current applications. For instance, a physician using AI to transcribe patient dialogues is best positioned to identify where the model excels or fails. These real-world limitations then serve as the roadmap for advanced research. To achieve this, interdisciplinary collaboration is essential.
