Q & A
5 minute read

All decisions have trade-offs. IBM’s Wei Sun is an expert at weighing them

The IBM researcher has applied theoretical concepts in operations research to finding a fair pricing policy for airline tickets and a cost-effective way to serve LLMs to IBM customers. We spoke to her about her OR, AI, and what’s next.

In business and in life, choices come with opportunity costs. It’s part of what can make decision making so excruciating. Will the costs and benefits of choice A outweigh the costs and benefits of B or C? Wei Sun is an expert at navigating “what if” scenarios like these. But when looking back at her own career, she’s surprised by some of the decisions that brought her to where she is now.

As a senior research scientist at IBM, Sun brings the mathematical rigor and pragmatism of operations research to some of the knottiest problems in business. Growing up in Shanghai and later Singapore, she had wanted to be a writer or an artist. But when it came time to pick a major in college, she made the safe choice: electrical and computer engineering. “As an immigrant, job security came first,” she said. “I thought ECE would open doors to a stable career.”

Sun moved to the US to pursue a master’s at MIT, and eventually, a PhD in operations research. She has been at IBM since 2012, often working with clients to resolve those difficult business questions. She was recently recognized with a prestigious award in her field by the Institute for Operations Research and Management Sciences (INFORMS). The prize, for revenue management and pricing practice, was for work helping an airline client price set a pricing policy for their premium seats. It had to be fair and transparent — as well as profitable.

The proposal Sun and her colleagues came up with combined causal machine learning and mixed-integer optimization, which increased revenue for the airline by 7%, and was projected to raise $100 million per year if rolled out domestically.

Sun currently works as a technical assistant (TA) for David Cox, IBM Research’s VP of AI models, and has contributed to research aimed at lowering the cost of LLM inferencing by routing queries to the smallest, most capable models, when possible. She lives in New York with her husband, who she met at MIT, and their two daughters, ages 5 and 8. We recently caught up with Sun to chat about her career path, OR, and where AI is headed.

What motivated you to study electrical and computer engineering as an undergrad?

Honestly, as an immigrant, job security came first. Engineering felt like a safe, practical choice. But over time, I grew to appreciate the problem-solving mindset and creative thinking it fosters.

What brought you to MIT and why Operations Research (OR)?

In my final year of undergrad, I stumbled upon a master’s program co-run by MIT and other local universities. I applied on a whim, and somehow got in. I had to pick a research area, and I chose OR mostly by process of elimination. I knew the least about it, but I liked the other options — like chemistry and biology — even less. It turned out to be one of the best decisions I’ve ever made, as MIT is home to arguably the world’s top OR program. I liked it so much I ended up staying for a PhD.

What is “OR” exactly?

When they ask me at the immigration desk, I call it “applied math” and that tends to get me waved through. More specifically, OR draws on probability, statistics, algorithms, and optimization to help people make better decisions. You can think of optimization as navigating within constraints while still finding the best possible outcome.

What was your path to IBM Research?

I spent six years at MIT studying a game theory concept called the “price of anarchy.” It measures how much efficiency is lost when individuals act in their self-interest rather than the common good. Over time, I realized I wanted to do more than prove theorems on paper. IBM Research stood out to me because of its strong legacy in OR. Ralph Gomory’s pioneering work on integer programming showed that mathematical elegance could drive practical impact.

weiSun-inline.jpg
In these notebook pages, Wei Sun explores two AI approaches for high-stakes business decisions. At left, an end-to-send prescriptive neural network learns optimal actions from observational data. Prediction and prescription are combined in a single training process to create a model with interpretable rules. The sequential approach at right, by contrast, predicts counterfactual outcomes first, followed by large-scale mixed-integer optimization to determine the best actions. This modular design offers greater flexibility for businesses with existing prediction tools.

How did you get into AI-related research?

It’s hard not to at IBM! That said, my path was pretty unconventional. I spent much of my first decade working with external clients through IBM Consulting on proof-of-concept projects. As a theorist by training, it felt like an AI bootcamp. Once I understood how to apply models, digging into how they worked came naturally with my math training. Clients often asked, “What can we do to achieve X?” — a causal question. Yet most AI pipelines rely on correlational models. That disconnect pushed me to explore causal decision-making, a topic I’ve been passionate about ever since.

What AI capabilities are you most excited about?

Anything that frees people from repetitive drudgery so we can focus on what makes us human: being creative, caring for others, and engaging in more meaningful, fulfilling experiences.

Why do you think LLM routing is so important?

You don’t need an expensive frontier model for every query. A smart router can match queries to models good enough for the task, saving time and money without sacrificing quality. A lot of existing work, including IBM’s cost-aware CARROT method, assumes access to a labeled dataset of model responses which is expensive to curate, especially as new models come out. What if you could train a router on raw deployment logs instead? Our causal routing method takes CARROT a step further by relaxing data requirements and grounding the approach in causal reasoning. Watsonx is currently implementing CARROT and causal routing may be the next step since it addresses deployment limitations like data efficiency and bias correction.

Where do you see AI heading?

The real challenge is making AI more predictable, reliable, and efficient, so it can truly help people in meaningful ways. I’ve seen a lot of promising ideas in this space, both at IBM and across the broader community.

What fills your days right now?

I’m currently the TA to David Cox. I spend time shadowing him, sitting in on technical meetings, and staying connected to ongoing projects and strategic discussions. I also contribute technically — through data crunching and running optimization and simulations — to help answer strategic questions or stress-test hypotheses. The role has helped me identify places in the generative AI pipeline where OR could contribute, such as data gathering, as well as model training and inferencing.

How did you get interested in technology?

My interest in computers and technology came relatively late. As a kid, I imagined becoming a writer or an artist — creativity and self-expression were always important to me. But I’ve always been curious by nature, and that curiosity has led me to where I am today.

I got my first computer in middle school and told my mom it was for homework, but mostly I used it to play The Sims and SimCity, watching avatars set their kitchens on fire and trying to fix traffic jams I accidentally created. Along the way, I learned about budgeting, managing limited resources, and trying to keep my virtual citizens alive. In hindsight, I can see how it prepared me for OR and working in tech.

What do you like doing outside of work?

I love being in nature, jogging, gardening, and taking long walks around the IBM Research Yorktown campus. I love music, especially jazz and bossa nova. I also really enjoy going to concerts: This year, I’m excited to see Pink Martini and Laufey.

What’s next for you?

Honestly, I’m not sure. As you might be able to tell, I don’t tend to plan that far ahead. I like to think I’m optimizing for a different objective function: maximize experiences, expand learning, build meaningful relationships, and have some fun along the way. I'm aiming to stay on the Pareto frontier of life.

Related posts