Research

Advised by Prof. Kristian Hammond. Developed LLM product that analyzes real-time audio conversations, detects relevancy and misconceptions, and provides targeted Socratic questions and material suggestions through RAG. Groupal aims to help students work together more effectively and build a deeper understanding in study sessions. The project’s goal is to create a virtual learning assistant that listens to real-time student discussions, detects misconceptions, and facilitates discussions through Socratic questioning techniques and relevant background knowledge retrieval. ...

Advised by Prof. Stephen Xia. Research LLMs for medical decision-making; developed method for learning joint embeddings across CXR, ECG, and EHR data for zero-shot classification of cardiovascular diseases, achieving >60% accuracy. Credits: Nan, Yueyuan.

A Human-Inspired Solution to LLM Memory Enhancement Authors: Terry Chen, Kaiwen Che, Matthew Song Abstract Despite advances in large language model (LLM) capability, their fundamental limitation of not being able to store context over long-lived interactions persists. In this paper, a novel human-inspired three-tiered memory architecture is presented that addresses these limitations through biomimetic design principles rooted in cognitive science. Our approach aligns the human working memory with the LLM context window, episodic memory with vector stores of experience-based knowledge, and semantic memory with structured knowledge triplets. ...

Situated Practice Systems: Improving and Scaling Coaching through LLMs Authors: Terry Chen, Allyson Lee Abstract Effective coaching in project-based learning environments is critical for developing students’ self-regulation skills, yet scaling high-quality coaching remains a challenge. This paper presents an LLM-enhanced coaching system designed to support project-based learning by helping connect peers struggling with the same regulation gap, and to help coaches by identifying regulation gaps and generating tailored practice suggestions. Our system integrates vector-based semantic matching with LLM-generated regulation gap categorizations for Context Assessment Plan (CAP) notes. Results demonstrate that our system effectively retrieves relevant coaching cases, reducing the cognitive burden on mentors while maintaining high-quality, context-aware feedback. ...