Intelligence in how priors shape learning

Humans observe sparse and ambiguous data, yet from it they develop rich internal models to make inferences about the world. Large language models (LLMs) can also generalize to novel tasks from just a few in-context examples. How is that possible? These intelligent agents, humans and LLMs alike, have good priors that guide their learning and generalization efficiently. I work on understanding the structure of priors: what makes a good inductive bias, and how inductive biases constrain/guide learning, leading to good or bad generalization.

Science can also be viewed as a learning process, constructing and updating priors for better future learning. What is the best form of knowledge that explains the world but also fits the structure of the human mind? How can we describe the space of all possible forms of knowledge, constrained by the structure of the human mind (a universal grammar of knowledge)?

I am currently a postdoc with James Evans at the University of Chicago, studying these questions in the context of AI for science and AI alignment.

Some of my studies (some with 5-minute videos walking through the main results):

1. How representational structure constrains online learning:
   • Metacognition in LLMs (NeurIPS 2025)
   • Neural constraints on in-context learning (COLM 2026)
     🎬 Click to watch the video
   • In-context superposition: working memory limit in LLMs (COLM 2026)
     🎬 Click to watch the video
   • Hypothesis generation and updating in LLMs (NeurIPS 2026)
     🎬 Click to watch the video
2. Models of human behavior:
   • Humans adapt to the environment online like policy gradient descent
   • Foundation model of human behavior
   • Using human think-aloud to predict behavior

The pronunciation of my name is “HWAH-doang SHAWNG”.

I am on the job market

I am looking for full-time Research Scientist and faculty positions in using AIs to model human behavior (including the scientific process) and AI alignment.

Want to collaborate or chat?

I’m always happy to collaborate. Whether you’re interested in my research or just want to chat about science, shoot me an email. If you’re at UChicago, even better. I’m usually at the Knowledge Lab (1155 E 60th St), Room 203. Feel free to just walk in.

Some of my writing

I also write about cognition, science, and the concepts we use to understand the world. If something interests you, feel free to email me. I’d like to hear what you think.

Some interesting facts:

• I enjoyed reading when I was young. My favorite writers are James Joyce, Milan Kundera, Jorge Borges, Franz Kafka, Dostoyevsky and Edgar Allan Poe.
• I chose this stupid username (sakimarquis) as a teenager. It came from Saki and Márquez.
• Yet I haven’t read much since high school. I suddenly lost my patience with long books.
• I am addicted to computer games, but only when an exam is approaching. Since I have few exams to take, I seldom play them now.
• I enjoy sick jokes and embarrassing short videos.
• Using a second language is painful for me, mostly because I can’t help being sarcastic but I can’t do it well in English.
• I am bad at calculating. I often mess up single-digit calculations, even with a pen and paper. This always makes me doubt myself as a researcher in computational neuroscience. (But large language models also fail at simple calculations, so I am not alone.)