Research Homepage
송민규
Department of Computer Science
Yonsei University
AI Systems · Machine Learning Systems · Large-Scale Intelligent Systems
I am an undergraduate student in Computer Science at Yonsei University.
My interests lie in AI systems, machine learning systems, and the structural behavior of large-scale intelligent systems. I am especially interested in how large AI systems behave when they are deployed as real execution systems involving memory, runtime, hardware, inference, and reliability constraints.
Rather than treating AI models only as algorithms, I try to understand them as systems: systems that store, compress, retrieve, schedule, fail, and scale. My current work and study focus on building the technical foundation needed to analyze these systems from both computational and mathematical perspectives.
I am also interested in statistical physics, information theory, and quantum information as mathematical languages for thinking about computation, inference, and complex systems.
I am currently working on independent research projects around large-scale AI systems, with a focus on memory, inference, reliability, and system-level behavior.
My work is motivated by a simple question:
How do large intelligent systems preserve, lose, compress, and reorganize information under real computational constraints?
Current directions include:
Accepted Poster · ICML 2026 Workshop on Failure Modes in Agentic AI (FAGEN)
Agentic AI · Memory Retention · Failure Modes · Graph Robustness · Prerequisite Eviction
David B. Kirk, Wen-mei W. Hwu
Vijay Janapa Reddi
Ian Goodfellow, Yoshua Bengio, Aaron Courville
Nigel Goldenfeld
Marc Mézard, Andrea Montanari
Roger Penrose
Michael A. Nielsen, Isaac L. Chuang
Inference serving runtime for large language models
Model implementations and ecosystem for modern deep learning systems
My work draws on ideas from:
systems, runtimes, parallel computing, and large-scale inference
deep learning, agentic AI, mechanistic understanding, and reliability
statistical physics, information theory, complex systems, renormalization, and quantum information
I value open-source software not only as a way to share code, but also as a form of research infrastructure.
Many of the systems I study, including modern machine learning frameworks and large-model serving runtimes, exist because researchers and engineers have built shared tools that others can inspect, extend, and improve. I hope to contribute to this ecosystem as I continue to develop my own technical foundation.
I am also interested in AI-assisted research as a new epistemic medium for scientific and engineering work. As modern research fields become increasingly dense with accumulated literature, AI systems may help researchers explore large conceptual spaces, connect distant ideas, and generate candidate directions that would be difficult to discover through purely linear search.
I do not view this as delegating knowledge creation to AI. The responsibility remains with the human researcher: to define the problem, judge relevance, verify claims, formalize arguments, run experiments, and transparently record how AI systems contributed to the work.
In this sense, Human-AI Co-Research is not a way to bypass research ethics, but a demand for stronger ones: transparency, reproducibility, and clear human accountability over AI-mediated exploration.