Runbao (Frank) Du
I am a 3rd-year undergrad studying Computational and Applied Mathematics (CAAM) with a minor in Astronomy and Astrophysics at the University of Chicago. My main interest lies in AI applications and the intersections of AI, ethics, security, and music.
I enjoy making music, writing poems, and sometimes typing competitively. I also enjoy bouldering, working out, playing ultimate, and, on rare occasions, running 400m. I love making plans — I make too many plans for everything in my life and they are probably hindering my productivity. I enjoy thinking about big topics and vague problems and pretending to have solutions for them even though I don't even have a solution to the second question on my math homework.
I am always looking for new opportunities.
Email runbao (at) uchicago (dot) edu
CV
Last updated: September 2026
Research
Open-source Music-generation Model: MuseV2 ongoing
Building a controllable text-to-music LLM from scratch with Fudan University's NLP Lab. Currently working on data preparation. Built a deterministic cascade/DAG sampler that generates structured song specs across 17 genre families / 99 sub-genres with coverage-balanced, conditionally-dependent fields
Adversarial Analysis of AI-Music Detection and the Limits of "Humanness" ongoing
Building a rigorous, de-confounded detector for AI-generated music (Suno and beyond), and using an adversarial generator–detector loop to ask which artifacts of "machine-made" music are removable surface tells vs. stubborn deep fingerprints. Mentored by Qile He from Fudan University. Supervised by Prof. Blase Ur.
The Effects of Value-Laden Persona Descriptions on LLM Value Inferences and Task Completion Nov 2025 - May 2026
Worked in the UChicago's SUPERGroup Lab to explore value-based reasoning in large language models.
Abstract
Humans increasingly delegate everyday tasks like editing emails and soliciting recommendations to large language models (LLMs). Because many decisions implicit in completing these tasks reflect human values, we wondered to what degree a short self-description of a user (persona) might impact how LLMs completed tasks. In our first experiment, we studied the values LLMs infer. The eight LLMs we tested inferred from these 33 personas 2,583 unique values, which we formed into 162 clusters. In our second experiment, we tested how these personas impacted how eight LLMs completed 33 tasks. We found that nearly any persona, even if seemingly irrelevant to the task, shifted how LLMs completed tasks. Personas relevant to a task caused an even larger shift. In other words, we found that even a few sentences that suggest particular values can cause an LLM to complete a task far differently, raising tensions between personalization and stereotyping.
UChicago Directed Reading Program Aug 2025 – Jan 2026
Learned fundamental concepts in algebraic geometry and applied them to study elliptic curves, using Silverman's The Arithmetic of Elliptic Curves as the main textbook. The ultimate goal is to understand and reproduce the proof of the elliptic curve primality test.
Build
brain2 — Constructing, maintaining, and "Agentalizing" knowledge Sept 2026 – present
A tool for quickly building, organizing, and visualizing personal knowledge bases — what we call second brains. Notes can be typed, recorded, or imported from local files, then split and embedded into MongoDB, clustered by topic, and laid out on a 3D map. Each second brain is then represented by an agent, and a four-role debate engine with mid-similarity cross-brain retrieval lets those agents exchange ideas and produce reasoned, source-traceable conclusions. The long-term goal is to connect many people's second brains into a network that can be embedded in companies and teams, so that skills and knowledge flow without barriers.
Built the first version at AIx Origin Hack 2026 in Shenzhen, where I served as team lead, originated the idea, and owned the frontend and data layers of a three-person team.
Selected for the final pitch out of 70+ teams and awarded the ModelScope Open-Source Contribution Award.
Audit Copilot — LLM Agent for Multi-track Audio DatasetMay 2026 - July 2026
Assisted in building Audit Copilot, an in-house LLM-agent tool built for Fudan University's Lab of Audio and Music Technology, aiming to curate a large-scale Chinese-pop multi-track dataset (vocals, stems, MIDI, mix-project files) for music generation models. It automates the multi-hour manual QC pipeline — file-naming, structure, format, alignment, arrangement checks — that reviewers previously ran by hand for every song, driving a 17-state SOP with an LLM agent that hands control back to a human whenever subjective judgment (mixing, tempo, structure) is needed. Mainly worked on redesigning the Phase-B human-in-the-loop SOP in the agent's system prompt and trimming workflow.
诗宇 Shiyu — An Interactive 3D Universe for PoetryMay 2026 - July 2026
A WebGL poetry universe in which themes are stars and poems are orbiting exoplanets, letting readers explore, curate and publish work inside a navigable 3D galaxy. Its central design challenge was fusing three normally separate disciplines into one coherent product: astrophysically plausible celestial simulation, the reading rhythm of poetry, and a visual aesthetic that had to satisfy both scientific credibility and literary feeling.
Award-winning in 2026 WaiTan Hackathon in multiple tracks.
Lyra: text-music chatbot Jan 2026 - May 2026
Built and deployed a full-stack AI music platform where users upload a song and chat with an LLM that "listens" through a custom Python audio-analysis pipeline and generates on-the-fly visual scenes via Claude tool-use orchestrating Google's Gemini image model.
Please contact Runbao for the prototype!
Prezuricle — Multimodal Marketing Platform Aug 2023 – May 2024
Co-created with Jason Zhao (USC '27). An AI-driven platform covering multimedia conversions like text-to-speech and text-to-image, with the main services being AI-centric marketing tools.
Music
Whatnot
Animal names I use to plan the sh*t out of everything
I started experimenting with making plans and time management since grade 8 (or maybe 9 — I can't remember). I wrote a book about the planning techniques I used to survive high school, which nobody cared about or read.
I realized that the most important aspect of a good plan is to be flexible and multidirectional. A good plan should incorporate things that you want and don't want to do, but it should be so flexible that you are not restricted. The plan should not be guiding you. You should be guiding the plan.
I tend to use these animal categories with the following rules:
- I should be completing at least 3 animals a day.
- I should be prioritizing these animals at the start of each day so I have a general sense of what's more pressing, but rule 1 still applies.
- Some animals may take longer than the rest, and that's fine.
- Some tasks may fall under multiple animals. That's fine — completing that task counts towards completing multiple animals.
- If a task does not fall under any of these categories, it is not a task.
A table
