From AI to artistry, Binghamton University sees a summer surge in research

About 40% of Binghamton University’s undergraduates complete some form of research before they graduate, according to Stephen Ortiz, the assistant vice provost for academic enrichment and the director of the External Scholarships and Undergraduate Research Center (ESURC). His office aims to increase that number to accommodate every student who wants to do research.

“A transformative learning community is not just curricular in its orientation but also has a lot of experiential learning,” Ortiz said. “We’re trying to get every student involved in one of these ‘high-impact practices’ — study abroad, internships, undergraduate research, community-engaged learning or service learning — by the time they graduate.”

In 2023-24, Binghamton University initiated a new effort to increase participation in faculty-mentored undergraduate research.

Finding faculty-mentored research opportunities can be a challenge for students. Enter the Binghamton University Projects for New Undergraduate Researchers pilot program.

“It’s hard to establish a personal relationship. Maybe you’re a first-generation student, for example, and you find that intimidating. But do a Google form that says, ‘I want to work with that person’ and do an interview, it’s very clean and easy. It’s a model that works,” Ortiz said. “We are hoping for that next step, which is to entrench it as an ongoing program.”

The Binghamton University Projects for New Undergraduate Researchers (BUPNUR) launched in the spring 2024 semester to pair students with no previous research experience with faculty-mentored projects. The 2024 Chancellor’s Summer Research Excellence Fund included an additional $250,000 in funding for students to do summer research who are first-generation, have demonstrated financial need or are transfer students.

This initiative included 52 faculty members (including 14 new faculty hired through the SUNY hiring initiative). These faculty hosted 38 projects across STEM and humanities fields; Students simply select their top projects and apply.

Ultimately, 70 undergraduates were awarded positions last spring after interviews — all of whom had no previous research experience and were provided stipends of $1,500-$2,500. Seventeen BUPNUR students will participate in the program this summer.

AI can be a DReaM

Eight of these undergraduates are working in artificial intelligence research, a particular goal of the Chancellor’s Summer Research Excellence Fund. These include four students of William Hayes, an assistant professor in the Department of Psychology.

Hayes leads the Decision Research and Modeling (DReaM) lab at Binghamton University. His work focuses on how context can shape choice behavior and cognitive processes.

His summer research project, though, focuses on “machine learning” and how it mimics human choice. As a new faculty member, Hayes was excited to offer his services to the project and explore a multidisciplinary approach — three students are computer science majors, while the fourth is majoring in psychology.

“I liked the idea of trying to involve people who had not been involved previously in a lab,” he said. “I liked the format of the project, and it was also a great opportunity for me to reach out to students in other departments, because the project is multidisciplinary.”

The group has split into two teams, though both are looking at large language model AI — such as ChatGPT — designed to predict what words come next given a pattern of text. The reason his field finds these models interesting is that some can do psychology tasks as a byproduct of their programming, like a human would.

“We’re asking the models to do something that they were not explicitly trained to do. It’s fascinating that they can do these tasks, but we still don’t have a great understanding of how,” Hayes said.

One of the group’s projects gives hypothetical decision problems to a model, each a choice between a safe option and a risky gamble. The two options always have the same expected value, but the team uses the model’s answer to gauge its risk appetite. This is a common task given to humans in decision-making labs.

Similarly, the second project explores a problem known as “multi-armed bandit tasks.” Participants are presented with buttons and given the option to select them as many times as they want. The trick is that not every button has the same outcome, and participants must discover which one has the highest average reward. These kinds of decisions involve trade-offs between exploration and exploitation. When do you start capitalizing on the most efficient path and stop searching?

Isaac Cohen, a senior in computer science working on his 4+1 degree, is part of this second project. He and his partner have given the large language model five options — making it even more important to explore. To do this, the pair generates scenarios, translates them into text prompts, and then, critically, answers the question for the model to fit a designated outcome.

The students’ goal is to “extract” the model’s internal representation of the problem. They try to understand whether the model detects the difference between greedily choosing an option for its early reward success versus exploring for a potential success down the line. At some point, this kind of data may help them steer the models toward more optimal decisions for a situation’s needs.

“We’ve replicated some previous studies and tried to see if the model can tell if it’s using a particular strategy. We proved that it seems to know the difference,” Cohen said. “Now, we’re trying to make it do one or the other, by steering — when you add a vector into the model while it runs, to influence it. It’s like the brain: you could put people through an fMRI and see what gets activated, and then you could try to change those activations for a particular behavior. That’s the next step for our machines.”

Although Cohen is currently researching large language models and finds the work interesting, he is not sure whether they represent the future of AI. It’s for this reason he thinks the focus on psychology in the DReaM lab is so important.

“Next year, people might find out that large language models are just not the way to do AI, and there’s a completely different way,” he said. “But the DReaM lab is fascinating, because it integrates decision-making in humans, and humans aren’t changing that quickly. The human mind is a lot more constant.”

Hayes agrees; his hope is for students like Cohen to see the similarities between cognitive science and computer science and learn how to apply an interdisciplinary approach in their future research to make it stronger. He hopes that the students will put together a research poster, and one day turn that into a paper to publish. Even without those goals, Cohen believes he earned valuable experience.

“For people like myself, projects like this benefit them greatly,” Cohen said. “Even if it turns out that there’s not much application, I’ve learned a lot about large language models this summer, just by tinkering with them. I’ve learned a lot in the process about code, and I learned how things work in academia and how research works in the real world.”

Ortiz hopes the project will be renewed in the new year, with potential funding coming from even more innovative sources. He believes that programs like BUPNUR make Binghamton a special kind of institution.

“Research is really the spirit of inquiry, of the collaborative collective. At Binghamton, there is a culture of supporting, communicating and valuing people asking hard questions while trying to find answers to them,” Ortiz said. “This development of a culture of undergraduate research on campus, that sees students not just as recipients of knowledge but creators of knowledge themselves, is an incredibly wonderful and vibrant addition to our campus.”