PeopleofSuperintelligence

Researcher

Michael Y Hu

Also known as Michael Hu

Michael Y Hu is a doctoral researcher at NYU's Center for Data Science who studies the science of training language models, with a focus on data. As a Princeton undergraduate, Hu worked with Karthik Narasimhan and Tom Griffiths, and in 2022 co-authored a paper on instilling human inductive biases in machines that NeurIPS named an Outstanding Paper. At NYU, they worked with Kyunghyun Cho and Tal Linzen on training dynamics, data mixing and downstream scaling laws, and their 2025 paper on pre-pretraining with formal languages won an Outstanding Paper Award at ACL 2025. In 2026, they introduced OP-Mix, a data mixing algorithm for the whole language model training lifecycle. They are a fifth-year PhD student advised by Cho and Linzen and a member of technical staff at Principia.

Now at
New York University and Principia

Timeline

  1. 2022

    A NeurIPS Outstanding Paper

    In 2022, Hu is one of ten authors of "Using natural language and program abstractions to instill human inductive biases in machines", a paper on meta-learning agents that includes Karthik Narasimhan and Tom Griffiths, the Princeton researchers Hu works with for two years during a BSE in computer science. NeurIPS names the paper an Outstanding Paper in 2022.

  2. 2023

    Hidden Markov models of training

    Hu is first author, in 2023, of "Latent State Models of Training Dynamics", written with Angelica Chen, Naomi Saphra and Kyunghyun Cho and accepted at TMLR 2023. The method fits a hidden Markov model over metrics recorded during neural network training and uses the latent states to study phase transitions and detour states that slow down convergence.

  3. 2024

    BabyLM findings and Aioli

    In 2024, Hu is first author of "Findings of the Second BabyLM Challenge", a report on 31 submissions to a competition to train language models on 100 million words or fewer. They also co-author Aioli with Mayee Chen, Nicholas Lourie, Kyunghyun Cho and Christopher Ré, an online data mixing method that estimates mixing law parameters during training.

  4. 2025

    An ACL Outstanding Paper

    Hu leads "Between Circuits and Chomsky: Pre-pretraining on Formal Languages Imparts Linguistic Biases" in 2025, written with Jackson Petty, Chuan Shi, William Merrill and Tal Linzen. Pre-pretraining on formal languages that capture hierarchical dependencies gives lower loss and better linguistic generalization on natural language, and a 1B-parameter model reaches the same loss with a 33% smaller token budget. ACL 2025 names the paper an Outstanding Paper.

  5. 2025

    A scaling laws reality check

    The 2025 paper "Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check", co-authored with Nicholas Lourie and Kyunghyun Cho, appears in EMNLP Findings 2025. Its meta-analysis of existing data finds predictable downstream scaling in 39% of cases.

  6. January 2026

    Neural Neural Scaling Laws

    In January 2026, Hu and four co-authors, Jane Pan, Ayush Rajesh Jhaveri, Nicholas Lourie and Kyunghyun Cho, post "Neural Neural Scaling Laws". Their NeuNeu network, trained on open-source model checkpoints, predicts accuracy on 66 downstream tasks with 1.99% mean absolute error, a 44% reduction against logistic scaling laws.

  7. As of October 2026

    Technical staff at Principia

    Hu is a fifth-year PhD student at NYU's Center for Data Science, advised by Kyunghyun Cho and Tal Linzen, and a member of technical staff at Principia. In May 2026, they are first author of "Always Learning, Always Mixing", which introduces OP-Mix, a data mixing algorithm that simulates candidate mixtures by interpolating low-rank adapters. Their research studies the science of training language models, with a focus on data, and their earlier career includes two years as the first employee of Yobi AI.

Career

  1. NowNew York UniversityDoctoral Researcher
  2. NowPrincipiaMember of Technical Staff
  3. Yobi AIFirst Employee

Education

  1. Princeton UniversityBSE in computer science
  2. New York UniversityPhD

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes New York University. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes New York University. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  3. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes New York University. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  4. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes New York University. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Michael Y Hu yet. Send us a free-licensed photo.

  1. Michael Hu michahu.github.io. Checked 21 October 2026.
  2. [2205.11558] Using Natural Language and Program Abstractions to Instill Human Inductive Biases in Machines arxiv.org. Checked 21 October 2026.
  3. Announcing the NeurIPS 2022 Awards – NeurIPS Blog blog.neurips.cc. Checked 21 October 2026.
  4. [2308.09543] Latent State Models of Training Dynamics arxiv.org. Checked 21 October 2026.
  5. [2412.05149] Findings of the Second BabyLM Challenge: Sample-Efficient Pretraining on Developmentally Plausible Corpora arxiv.org. Checked 21 October 2026.
  6. [2411.05735] Aioli: A Unified Optimization Framework for Language Model Data Mixing arxiv.org. Checked 21 October 2026.
  7. [2502.19249] Between Circuits and Chomsky: Pre-pretraining on Formal Languages Imparts Linguistic Biases arxiv.org. Checked 21 October 2026.
  8. Awards - ACL 2025 2025.aclweb.org. Checked 21 October 2026.
  9. [2507.00885] Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check arxiv.org. Checked 21 October 2026.
  10. [2601.19831] Neural Neural Scaling Laws arxiv.org. Checked 21 October 2026.
  11. [2605.15220] Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time arxiv.org. Checked 21 October 2026.