Michael Y. Li

Hi! I'm a Computer Science PhD student at Stanford, where I'm advised by Noah Goodman and Emily Fox.

Previously, I graduated summa cum laude and Phi Beta Kappa from Princeton, where I was fortunate to work with Tom Griffiths and Ryan Adams.

Research

I work broadly on reasoning in LLMs. My previous work explores principled, statistically rigorous ways to integrate black-box LLMs into data-science workflows—for example, how we might leverage these models for experimental design, hypothesis testing, or statistical modeling. Before that, I worked on statistical machine learning (e.g., Sequential Monte Carlo, variational inference, Gaussian processes, point processes). My email is firstname.middle_initial.lastname@stanford.edu!

Automated Hypothesis Validation with Agentic Sequential Falsifications

LLMs + sequential hypothesis tests with Type-I error control

BoxingGym: Benchmarking Progress in Automated Experimental Design and Model Discovery

A benchmark for LLM driven experimental design and model discovery

CriticAL: Model Criticism Automation with Language Models

Hypothesis tests + LLMs

What Should Embeddings Embed? Transformers Represent Latent Generating Distributions

We study the embeddings of transformers through the lens of predictive sufficient statistics.

Automated Statistical Model Discovery with Language Models

We propose a language model driven automated statistical model discovery system.

NAS-X: Neural Adaptive Smoothing via Twisting

We introduce a new method for inference and model learning that combines reweighted-wake sleep and smoothing Sequential Monte Carlo. We theoretically analyze the bias and consistency of our method and then apply it to discrete latent variable modeling and fitting mechanistic models of neural dynamics.

Why think step-by-step? Reasoning emerges from the locality of experience

We empirically and theoretically study when chain-of-thought reasoning emerges in large language models.

Gaussian Process Surrogate Models for Neural Networks

We propose a framework that uses Gaussian processes to approximate neural networks. We use this framework to analyze neural network training dynamics and identify influential data points.

Learning to Learn Functions

We propose hierarchical Bayesian models of how people learn to learn functions and validate our model in behavioral experiments.

Design and source code from Jon Barron's website