The latest interview in our series with the AAAI/SIGAI Doctoral Consortium participants features Ximing Wen who is researching transparent and trustworthy AI systems. We found out more about her work, her experience as a research intern, and what inspired her to study AI.
Tell us a bit about your PhD – where are you studying, and what is the topic of your research?
I’m a PhD candidate in Information Science at Drexel University in Philadelphia. My research is about making AI systems more transparent and trustworthy. Right now, language models can give you very confident-sounding answers, but there’s no easy way to check whether those answers are actually correct or where they came from. I’m working on building models that can show their reasoning and point to the evidence behind their outputs — so that people can actually trust what the AI tells them, especially in areas like healthcare and legal document review.
Illustration of prototype architecture for text classification.
Could you give us an overview of the research you’ve carried out so far during your PhD?
My PhD started with a question: can we make interpretable models that are actually good enough to use in practice? Previous interpretable models always fell behind black-box models in accuracy, which made them hard to adopt in practice. I developed a prototype-based approach that closes that gap — the model explains its decisions by showing you similar examples it learned from, without losing performance. From there, I extended this idea to generative models — exploring whether a model can not only give you an answer, but also show you exactly where it found that answer. Along the way, I’ve also applied these ideas to medical AI, building interpretable diagnostic tools that work even with very limited training data.
Screenshot of our iOS application running pneumothorax (PTX) diagnosis on a 12.9 inch Apple iPad Pro.
Is there an aspect of your research that has been particularly interesting?
Definitely the spatial grounding work. There was a moment where I redesigned how the model learns about spatial coordinates, and accuracy jumped from around 65% to over 85%. The original loss function was basically blind to small regions in documents, so the model just learned to ignore them. Once I introduced a scale-aware loss, everything changed. It was a powerful reminder that how you teach a model matters just as much as the model itself — which is really the core idea behind my entire dissertation.
What are your plans for building on your research so far during the PhD – what aspects will you be investigating next?
So far, my research has shown that we can make language models more transparent — through prototype-based reasoning for classification and spatial grounding for document question answering — without giving up performance. The natural next step is to push these ideas further. I want to bring prototype-based interpretability to larger generative models. Most prototype-based methods today only work for classification — extending this kind of case-based reasoning to generative models is still a big open challenge. One direction I’m exploring is analyzing how different layers of a model encode different types of knowledge, and using that structure to build richer, more fine-grained explanations for model outputs. I’m also exploring how to make the AI alignment process itself more transparent, by integrating prototype-based reasoning into reward models. The idea is that if we can explain why a reward model prefers one response over another, we can build safer and more trustworthy AI systems.
Could you tell us about your research experience at Samsung and Amazon?
At Samsung Research America in Mountain View, I worked as an NLP research intern on the Language Intelligence team. I tackled a problem that sounds simple but turned out to be really challenging: can an AI read a complex document, answer a question about it, and point to exactly where it found the answer? Think of a doctor reviewing a 20-page medical report — they don’t just want the AI to say “the patient has condition X,” they want to see the exact paragraph that supports that conclusion. I developed new training methods that teach the model to understand spatial relationships between coordinates in a document, which dramatically improved its ability to locate answers accurately. That work was accepted at ACL this year.
During a summer internship at Amazon as an applied scientist, I worked on a different but related problem: making AI outputs that people can actually understand and trust. Amazon’s marketplace has millions of products across thousands of categories, and each category needs a clear definition that accurately covers everything within it. Previously, these definitions were written manually — a process that took weeks and still couldn’t keep up with the pace of new products and emerging categories. I built a system that generates these definitions automatically, and it actually outperformed human-crafted definitions in both accuracy and clarity. To me, that was a compelling example of AI’s potential: when the task involves synthesizing information across millions of items, AI can actually produce more accurate and consistent results than manual effort — as long as the outputs are designed to be clear and trustworthy.
Looking back, both experiences reinforced the same lesson: building powerful AI isn’t enough. If people can’t understand or verify what the model is telling them, the technology doesn’t reach its full potential.
What made you want to study AI?
My journey started at the end of my undergraduate studies, when I was working on my final year project. I trained a simple neural network on MNIST — a basic handwritten digit dataset — and was amazed that such a small model could achieve over 95% accuracy. That moment sparked something in me: if a simple network can understand an image, can it understand human language? Can it have real conversations with people? I became genuinely passionate about pursuing that question. And with the advent of GPT and large language models, much of what once felt like science fiction has actually come true. But the more powerful these systems become, the more I find myself asking a new question: can we make them safe and trustworthy enough for people to truly rely on? I believe that’s the key to AI reaching its full potential — and that’s what drives my research today.
And finally, what do you enjoy doing outside of the PhD?
I try to get outside as much as possible. I love taking walks along the river in Philadelphia’s Fairmount Park and watching the sunset after a day’s work, hitting the slopes in the Pocono Mountains in winter, or just kayaking and floating on the river in summer. Nature is the best way to recharge — it keeps me balanced both mentally and physically.
About Ximing
Ximing Wen is a PhD candidate in Information Science at Drexel University, where her research focuses on making language models more interpretable and trustworthy. Her work spans prototype-based interpretable models for text classification and spatially grounded architectures for document question answering. She has published at venues including ACL, COLING, and AAAI, and has contributed to federally funded research projects supported by NIH and DARPA. She has also gained industry research experience at Samsung Research America, Amazon, and Ping A Technology. Outside of research, she enjoys kayaking, skiing, and exploring nature around Philadelphia.