Shengtao "Alex" Ding

I am a high school student at Worcester Academy (Class of 2027) with a deep passion for Mathematics and Computer Science. My academic journey is defined by rigorous research and competitive problem-solving.

Currently, I am an Independent Researcher at MIT PRIMES, working on the mathematical foundations of topic modeling and NLP. I have also conducted research on causal discovery algorithms and on applications of group theory.

Beyond research, I led the Worcester Academy Math Team to regional victories and founded the Finance and Investment Club. I am an active participant in national math competitions, including ARML, HMMT, and HiMCM.

I am also currently serving as Worcester Academy's International Board of Monitor, Head Ambassador, and the founder of the Finance and Investment Club. If you're an international student interested in learning more about WA, feel free to reach out.

1.Do bigger language models actually produce clearer, more coherent topics, or can a lightweight encoder do the job just as well? In this study, we benchmark seven transformer embedding models inside a BERTopic pipeline across 11 datasets and find that scaling from ~22M to 13B parameters barely changes topic coherence or diversity. Click in to see the results, key tables, and what they mean for building faster, cheaper topic models without sacrificing quality. (Click on 'PRIMES_Final_Paper')

My research interests focus on the intersection of Mathematics and Computer Science, particularly:

• Mathematical Foundations: Exploring fundamental problems in abstract algebra, geometry, and topology.
• NLP Topic Modeling: Developing new algorithms for pattern recognition and classification in text data.
• Fractal Geometry: Studying mathematical properties of complex structures and their applications in natural phenomena.
• Causal Discovery Algorithms: Building mathematical models to infer causal relationships from observational data.
• Group Theory Applications: Applying symmetry analysis to molecular structures and cryptography problems.

These interests drive my participation in the MIT PRIMES research program, where I collaborate with mentors to explore cutting-edge applications of mathematical theory.

My research project uses clustering algorithms to analyze macro-level, trend-driven behaviors shaped by human decision-making, such as political prediction questions like "Who will ultimately become Trump's Secretary of the Treasury?" I analyze behaviors like betting, trading, and capital flows to turn these signals into predictive insights.

This is an interactive Trader Analytics web dashboard that analyzes 825 randomly selected prediction-market traders from Polymarket using K-means clustering, PCA, and network community detection (Louvain). The site includes a dashboard, a PCA cluster map, profit concentration (Lorenz/Gini), z-score cluster heatmaps, a trader network graph, leaderboards, trader lookup, and a cluster-to-community flow diagram.

What the Data Says: 3 Trader Archetypes

Using 20+ trading metrics (snapshot collected in Nov 2024), the clustering finds 3 clear segments with a silhouette score of 0.472 (reasonable separation).

If you're interested in my clustering analysis and the interactive visualizations I built, please click the button below for details: