Michael Y. Li

Hi, I'm Michael! I'm a CS PhD student at Stanford, advised by Noah Goodman and Emily Fox.

Previously, I graduated from Princeton (summa cum laude, Phi Beta Kappa), where I worked with Tom Griffiths and Ryan Adams. I’ve also spent time in tech and quant research.

X  /  Email  /  Google Scholar  /  LinkedIn

Research

I'm broadly interested in reasoning in language models, with a recent focus on making test-time computation more resource-efficient. Can reasoning models learn to garbage-collect their own KV cache? Some of my work draws on ideas from statistics—for example, can classic Monte Carlo methods help scale parallel test-time compute more efficiently? Previously, I worked on statistical machine learning, including twisted Sequential Monte Carlo. If you're a Stanford student interested in collaborating, feel free to reach out at firstname.middle_initial.lastname@stanford.edu! I've collaborated with a number of great undergrads on a few projects.

Selected Publications

Neural Garbage Collection

QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling

SPIRAL: Learning to Search and Aggregate

Simplified Sparse Attention via Gist Tokens

NAS-X: Neural Adaptive Smoothing via Twisting

All Publications

2026

QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling

auto-psych: Automating the science of mind using agent-driven theory discovery and experimentation

SPIRAL: Learning to Search and Aggregate

Simplified Sparse Attention via Gist Tokens

Neural Garbage Collection

GIANTS: Generative Insight Anticipation from Scientific Literature

2025

Simple, Scalable Reasoning via Iterated Summarization

We study iterated summarization—alternating between summarizing long reasoning traces and reasoning over those summaries—as a simple, scalable way to extend test-time compute.

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

2024

CriticAL: Model Criticism Automation with Language Models

Automated Bayesian model criticism.

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.

2023

NAS-X: Neural Adaptive Smoothing via Twisting

Twisted Sequential Monte Carlo for inference in latent variables models.

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

Using Gaussian processes to approximate neural networks.

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