About Me

Hello! My name is Roger Chen, and I am currently a rising senior (class of 2027) at Duke University with a double major in Mathematics and Computer Science. I've lived in Nashville, Tennessee all my life. My favorite hobbies are basketball, playing board games like chess, and reading. I also listen to music constantly, most notably nostalgic songs or classical music. I'm excited to showcase the work I've been doing at the DIMACS REU program for 2026.

Project Description

I am working on "Truth Learning in a Social Network" under the mentorship of Rutgers Professor Dr. Jie Gao along with collaborators John Wu and Eli Yablon. Essentially, we are analyzing how a group of agents may make decisions based on their own (potentially erroneous) information as well as limited observations of other agents' decisions. These situations can often create phenomena of information cascades or herding, which may prevent learning optimally as a whole even when there are theoretically infinite opportunities to gain information.

Weekly Log

Week 1 (May 27-May 30)

I arrived at the DIMACS REU program and attended orientation on Wednesday, a website workshop on Thursday, and two invited guest talks on Friday. My collaborators and I also met with Dr. Gao on Wednesday to review the document of project information that she had sent us beforehand and ask any preliminary questions.

Week 2 (May 31-June 6)

My group met with Dr. Gao at the beginning of this week to discuss our ideas for what we wanted to work on this summer (i.e., how we would build on the work that previous projects have accomplished for this project). Within the DIMACS program, we participated in project introductory presentations on Tuesday (see our slides in Resources). Once we had our research goals established, my group spent this week searching for past papers that had achieved similar objectives and tried to imitate related computations on our own. We again met with Dr. Gao on Friday, who listened to our progress and brought up new directions to consider.

This week, I also attended several talks at the DIMACS Computational Geometry Week (which Dr. Gao also helped organize). My favorite talk was by Wednesday's invited speaker (Dr. Suresh Venkatasubramanian) on bringing geometric intuition to the ethics of AI. For instance, he discussed how going from acceptable to unacceptable answers within a language model could be represented by a vector in n-dimensional space (this is the same for any other concept such as a gender change from male to female), which could be exploited with consequences.

Week 3 (June 7-June 13)

I started focusing on a more specific component of the overall research topic we are investigating, namely multi-pass learning. While this can be applied to any sort of network, I decided to focus on the family of cycle graphs, as inspired by my group's meeting with our advisor last week. From doing several computations on examples to determine if agents could converge to the majority of their private signals, I made many interesting observations and came up with various future directions related to NP-Hardness (relating to my collaborators' work) or general principles on convergence.

On Thursday, my group met with our advisor to discuss the progress that each of us had made. As this was a rather unfamiliar topic I was exploring, Dr. Gao helped by noting terminology that I had not known of (including "multi-pass learning"), which may help me delve into any past literature that may have worked on this topic before. We also better clarified my research question to be "what is the learning rate as number of vertices approaches infinity." In addition, she helped me take notice of general ideas already established in multi-pass learning, including the fact that convergence (towards the truth) can not be relied on to occur in cases of the complete graph/network.

Week 4 (June 14-June 20)

My group met with Dr. Gao on Monday. After I had spent time updating terminology and making my writeup on multi-pass learning more formal and comprehensive, this meeting was used to give me more ideas on specific directions that I could take within multi-pass learning, including how high-quality signals can propagate, the situation of binary trees, how multi-pass may be worse for majority vote, and how simulations can play a hand. Afterwards, I disproved a few rough "conjectures" I had had about general principles during multi-pass learning. (Unfortunately, these conjectures being true would have made my overall job a lot easier.) I also found a specific type of instance (in cycle graphs) where multi-pass can perform worse than single-pass, at least when we fix the ordering and private signals. This feels a bit forced, so more work should be done to see if we can see the same patterns in scenarios with more general parameters (e.g., no private signal-fixing).

Week 5 (June 21-June 27)

This week started with two separate meetings. First, the ACTION Institute, which helped fund our research project, hosted a Zoom meeting where I was introduced to members of the summer research program at UC Santa Barbara and very briefly gave overviews of our research projects. Our group also had our usual meeting with Dr. Gao, where we primarily discussed a new research direction that Eli thought of relating to agents' incentives to make decisions, as well as more multi-pass proofs in different parameters and settings.

John and I then spent the second half of this week traveling to the ACTION Institute at UC Santa Barbara to participate in their Year 3 site visit. There, I got to get a first-hand experience of how institutes like ACTION work with NSF and the government to gather sustained funding, and how the institute works as a whole to bring diverse workers towards the same overarching goals. It was also great to network with the ACTION interns at UC Santa Barbara, as well as several other researchers that I could look up to in mathematics, CS, and related fields.

Resources

[Project introductory presentation slides to come.]

Acknowledgements

This research is being conducted as part of the 2026 DIMACS REU program under the mentorship of Professor Jie Gao. My collaborators and I'd like to thank the NSF for supporting the DIMACS REU program under grand CCF-2447342. We would also like to thank ACTION (AI Institute for Agent-based Cyber Threat Intelligence and Operation), funded by NSF IIS-2229876.