Computer Science Department

Director: Vince Lin

Email: yivince32@gmail.com

Technical reports, algorithms, and machine learning research from our computer science program.

Artificial Intelligence in Computer Science

Computer science research exploring algorithms, model architectures, and systems. Click into papers to view figures (PNGs).

Compiling Human Thought into Precise LLM Prompts Workshop

Brooke Regan — 2026

As Artificial Intelligence integrates into STEM fields, technical efficacy relies heavily on natural language precision. “The Human Algorithm” is a 60-minute pedagogical workshop that simulates prompt engineering through a physical, active learning framework to expose the systematic pitfalls of human-to machine communication. Grounded in the principle of “Language as Code,” the curriculum demonstrates that logical execution depends entirely on structural syntax. Participants are divided into asymmetric roles: a “Visionary” (prompt engineer) who translates hidden geometric models into written instructions, and “Executors” (processing system) who replicate them without visual feedback. Across consecutive rounds, difficulty scales via constraints in orientation, scale, and shades to mirror system variables. The resulting visual discrepancies map the “Prompt Gap,” allowing educators to train students in debugging communication errors like structural ambiguity, logic gaps, and directional drift.

View PNG

Teaching Transformers with MicroGPT

Vince Lin — 2026

Slideshow introducing the transformer architecture, showcasing Andrej Karpathy's MicroGPT as an accessible tool for providing hands-on experience with tuning hyperparameters and understanding the structure behind modern Large Language Models.

View PNG

AI Coding Assistants: Revolution or Risk?

Angela Zhao — 2026

My slideshow briefly introduces the concept of ai coding assistants and discuss its advantages and disadvantages, as well as its influence on software development.

View PNG

Understanding Chain-of-Thought Prompting in Large Language Models

Cade Konishi — 2026

Chain-of-thought (CoT) prompting is a prompting technique which improves the performance of large language models (LLMs) by having them generate intermediate natural language reasoning steps before producing a final output. Chain-of-thought prompting is typically categorized into zero-shot CoT, where instructions are added to prompts, and few-shot CoT, where demonstrations with reasoning examples are added to prompts. CoT has been shown to improve performance for LLMs on tasks such as mathematical and commonsense reasoning, leading to state-of-the-art performance in certain benchmarks in these fields. Another upside of CoT is that it is very flexible in its applicability, as it can be applied to any task which can be reasoned through via language. However, CoT prompting comes with various limitations: CoT often generates incorrect reasoning paths, uses more tokens resulting in increased computational cost, and is less effective in smaller models. Overall, CoT prompting is a vital technique that greatly improves LLMs ability to perform complex reasoning.

View PNG

Exact Combinatorics vs Monte Carlo Tree Search in 9-Player AI Decision Systems

Hanyu Zhang — 2026

In games where 9 players compete and only one can win, the starting cards you get make a huge difference in your chances. But finding out exactly how much better the best starting cards are is hard because there are millions of possible outcomes. This research tests whether we can use C(n,r) to calculate the exact win percentage for the best starting position. We will also run a Monte Carlo Tree Search simulation to see if it gets the same answer.

View PNG

Internal Representations Within Transformer Architectures

Vince Lin — 2026

APA-formatted analytical paper on the paper "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task" (Li et al. 2023). The article seeks to summarize the findings outlined in the research paper regarding the internal representations of transformer models. By training a transformer model on Othello game sequences, the authors reveal novel insights into hidden-layer representations translating to human-like understanding after employing mechanistic interpretability techniques. These findings present promising conclusions on the application of board games in analyzing the properties of machine learning architectures.

View PNG