PhD students learn to bridge OR and AI at AI-SCORE conference

• Refine images for ship design 

• Create small therapeutic proteins 

• Accelerate materials discovery by analyzing crystal structure through deep reasoning networks 

• Identify metabolites using mass spectrometry 

• Aid in biostatistics by mapping species distribution and land cover in the Amazon 

Still, Gomes emphasizes that AI is just one resource for solving problems. 

“Yes, AI is great,” she says, “but it’s not going to get us there for scientific discovery.” Rather, she talks about the need for human scientists “to inject knowledge-centric reasoning into data-driven AI.” 

The jagged frontier 

Léonard Boussioux, an assistant professor of information systems at the University of Washington and another AI-SCORE presenter, describes himself as a vibe coding addict who uses generative AI regularly. He has ChatGPT review all his academic papers before he submits them and has sprung for the pro version of the tool. 

As hands-on work is an important part of the conference, Boussioux takes students deeper inside the AI itself, demonstrating how prompts are broken into tokens. He explains that, because AI “learns” from previous inputs, more of which are in English than other languages, using English as much as possible will yield better results.  

Boussioux demonstrates how he uses AI as a “co-iterator” through a prompt: “I want you to build a fun snake game in the theme of Chicago. I want this to be very fun and particular and original and creative.” He gives the same prompt to ChatGPT, Gemini, and Claude, using voice prompts for all three, “because it’s faster and you can be more spontaneous.”  

“This is how I build my understanding of the ‘jagged frontier’ of AI,” he says. 

After reviewing the games the AI had built, he asks students to build websites with Lovable, an AI-powered vibe-coding platform. 

Opening minds 

Equally as important as the faculty perspectives was the opportunity for students to learn from one another. 

During his talk, Fioretto puts Jinhao Liang, one of his students at UVA, on the spot to present on an optimization program their group had developed to create collision-free paths for robotic applications.  

But there are less-formal opportunities for learning as well. Chris Raymond-Bertrand, BS '25 from the Department of Industrial and Enterprise Systems Engineering at Illinois, who just finished the first year of his PhD program in industrial engineering at Virginia Tech, works more in the fields of probability and statistics than machine learning and neural networks. 

“Talking to others who are broadly interested in the same things, but whose approaches have been really different to mine, really helped open my mind,” he says.  

 He also appreciates the opportunity to hear from students and researchers who are working directly with AI, as does Allen Minch, who just finished his first year in Grainger’s PhD program in industrial engineering and operations research.  

“I don’t know that I necessarily see it as changing what I do for my research,” Minch says, “but I think it’s still very good to start to be informed … this session, my eyes have especially been opened with AI. I’ve come to realize it’s capable of a lot of things I didn’t realize it was capable of.”   

And, of course, there’s simply the opportunity to socialize.  

“That’s the best part,” says Zhengtao Su, who just finished his first year in Grainger’s PhD program for industrial engineering. “You get to know other PhD students from other programs, and you can communicate with them, learn from them.” 

 Other highlights during the week 

• Keynote addresses from Cornell’s David Shmoys and University of Pennsylvania’s Hamsa Bastani 

• Sightseeing in downtown Chicago, including cruises on the Chicago River and Lake Michigan and a tour of the Art Institute  

• Tutorials from the University of Southern California’s Vishal Gupta and Georgia Tech’s Kai Wang on contextual optimization and decision-focused learning. These approaches merge traditionally siloed areas, upending the assumptions that estimation is a machine-learning problem and decision-making is an optimization problem. In that combination, researchers create a seamless end-to-end process that combines both techniques. Like the GenAI sessions, these presentations had a significant hands-on component. 

The bottom line

Early in her remarks Gomes shows the students a “research subway map” she’d created to illustrate the themes and connections between her areas of work. 

Impressed, Boussioux decides to create his own subway map over the lunch break. He uploads his CV to Claude, along with a screenshot of Gomes’s map, and showed the results to the students. Gomes suggests he make an admin dashboard to allow him to make some adjustments himself. 

He loves that idea.  

“Carla,” he says, “you are an AI native vibe-coder.” 

 Using the dashboard, Boussioux makes a few revisions to his map, but he isn’t quite satisfied. 

 “Your map looked so much better than mine,” he tells Gomes, “because your map was created by a human.” Although Boussioux is an enthusiastic user of multiple AI platforms, his overall take on the technology is nuanced. Like everyone at AI-SCORE, he’s aware of the lurking fear that AI will come for everyone’s jobs. 

 He tells the students, “This is where your intuition — seeing things no one else can see — is precious and unique. AI won’t take that away from you.”