Chupei Wang

Researcher on LLM memory & interpretability

About

With a background in physics (University of Virginia) and philosophy—including a two‑year gap during college at the University of Chicago Divinity School (Eastern philosophy of religion)—I explore where logic and mind meet their limits, probing how the edges of science and the humanities intersect. I am driven by a curiosity about where cognitive architectures—biological and artificial—break down, and what these failures teach us about intelligence.

More about me

I received a bachelor’s degree in Physics from the University of Virginia, where I was particularly interested in how relativity helped transform the paradigm of physics. I was later admitted to a graduate program in philosophy of physics at the University of Bristol but did not enroll, choosing instead to conduct independent research in LLM interpretability. During college, I took several gap years and spent two years at the University of Chicago Divinity School studying comparative philosophy of religion, particularly ancient Eastern religious thought and its relationship to modern philosophy.

I like debating ideas internally and searching for connections between phenomena that appear unrelated across science and the humanities.

I also have startup experience in media production, working on music supervision and audio post‑production.

I welcome discussion and collaboration — email me.

Why approach LLMs from a cognitive‑science angle?

After earning my undergraduate degree in physics, I spent the past two years self‑studying philosophy of science (admitted but not enrolled to a philosophy program in the University of Bristol). During college, I also took a two‑year gap at the University of Chicago Divinity School to study the philosophy of religion (mainly Eastern). I was interested in how humans think and formulate theories from physics to philosophy. All those experiences inspired me to independently initiate and fund this PI‑for‑LLMs research. The work was accepted to the ICML 2025 Long Context Foundation Models Workshop, selected for an oral presentation at the AAAI 2026 LaMAS Workshop, and accepted to COLM 2026 Main. This recognition boosted my confidence and made me eager to learn more and continue thinking and observing.

I like making cross‑disciplinary comparisons and standing and observing at the intersections between the human and the natural world—from the mind and cognitive mechanisms to physics and algorithms. In particular, I am interested in uncovering where the structural features of knowledge within a field shape both its unique insights and its blind spots. The first cognitive‑with‑LLM project also grew from a gap‑year startup PM experience: with human engineers, spec clarification was natural; with LLMs, conversations felt different—fluent code but gaps in context‑aware reasoning. In Jan 2025 I began probing working memory; a proactive interference (PI) paradigm exposes this bottleneck cleanly.

Projects

Proactive Interference (PI) in LLMs — Cog4LLM

Interference-limited retrieval, measured minimally; simple mind‑flow with optional details.

COLM 2026 Main AAAI 2026 LaMAS Workshop — Oral ICML 2025 LCFM Workshop Ongoing Boundary cases Population codes

Open Demo Details / Progress

Contributions (CW)
  • CW (Chupei Wang): Initiated cognitive‑inspired WM stress tests; framed LLM retrieval errors as interference; led iterative trial‑and‑error testing; designed Experiment 1; pioneered prompt‑hacking interventions.

Paper: arXiv · Code: GitHub

Interpretability: Population Codes → Circuits → Boundary Cases

Taking cues from population coding and inhibition/gating motifs in the brain, I use circuit analysis to ask whether similar population-like representations and control signals exist in LLMs—and where they break. Early evidence suggests limits that appear in models but not in human data. I focus on boundary cases: stress tests that expose failure surfaces, then map them back to candidate mechanisms (ranking, selection, inhibition) to turn cognitive hypotheses into falsifiable circuit predictions.

Ongoing Boundary cases Circuit analysis

Ongoing — progress posted weekly.

Open project

Creative — Music Supervision (The Three‑Body)

Supervised music for early experimental animation series on Bilibili; coordinated NYU composers; oversaw post‑production.

Music Animation

Watch clip

更多 / More
  • 协调纽约大学作曲系成员，完成《三体》第一代实验动画（版权方实验系列）的配乐沟通与监督；系列在 B 站累计播放约 1.2 亿+。
  • 自第二季第 3 话起担任音乐监制；示例片段署名约 24:50 处（音乐制作人员名单：监制）。
  • 职责：以文字与示例沟通作曲意图与改进意见，统筹后期音乐制作。

Humanities & Philosophy — Zhuangzi and Tiantai

Seminar work at UChicago Divinity School with Prof. Brook Ziporyn, on classical Chinese (Zhuangzi; Tiantai). Minimal note with reference links.

Humanities

Open project

Contact

Email (main): wangchupei@hotmail.com

UVA (secondary): cw4bb@virginia.edu

Download CV