Core Francisco Park

corefranciscopark@g.harvard.edu

Cambridge, MA, USA

Hi! I’m Core Francisco Park (pronounced Corae), I am a 6^th year graduate student at Harvard Physics.

Currently, I work on understanding AI systems using carefully designed experiments. For the last few month I’ve been running experiments with synthetic data to understand compositional generalization and in-context learning. These days I’m mostly interested in understanding fine tuning and the long-context limit

In the past, I worked on developing reliable and deployable ML tools and methods for the physical sciences. I worked on different real data problems with ML and statistics, ranging from studies of a brain smaller than a grain of salt to cosmic volumes and from theoretical studies to practical ones.

Feel free to ask me anything!

-Core

latest news

Dec 10, 2024	I will be presenting some of our works at NeurIPS 2024: Our work on concept learning and hidden emergence of compositional capabilties will be presented as a spotlight poster! Our work on algorithmic phases of in-context learning will be presented as a poster at the Workshop on Scientific Methods for Understanding Deep Learning. Our work on how LLMs can reorganize their representations in a context dependent way will be presented as a poster at the Workshop on Symmetry and Geometry in Neural Representations.
Dec 01, 2024	Our paper “Competition Dynamics Shape Algorithmic Phases of In-Context Learning” is on ArXiv!
Oct 10, 2024	Our paper “Dynamics of Concept Learning and Compositional Generalization” is on ArXiv!
Sep 27, 2024	Our paper: “Emergence of Hidden Capabilities: Exploring Learning Dynamics in Concept Space” is accepted to NeurIPS 2024 as a Spotlight!
Jul 03, 2024	Tunadorable made a video about our paper: “Emergence of Hidden Capabilities: Exploring Learning Dynamics in Concept Space”