Miaosen Chai

I'm an incoming PhD student at UChicagoUniversity of Chicago, advised by Prof. Chenhao Tan. I'm interested in the science of foundation model training and building AI Scientist system that trains the next generation of AI by democratizing the scientific discovery process.

Previously, I received my B.S. in Computer Science from USCUSC, where I had a wonderful time as an undergraduate researcher advised by Prof. Robin Jia, Prof. Willie Neiswanger, and Prof. Jesse Thomason.

In industry, I spent some time as a researcher at ORNLOak Ridge National Lab and AI2AI2, and as an investment fellow at HongShanHongShan (formerly Sequoia China).

Feel free to reach out if you want to chat!

Contact Info  /  Google Scholar  /  Twitter  /  Misc

profile photo

News

  • 2026-09 Precise Debugging Benchmark: Is Your Model Debugging or Regenerating? accepted to NeurIPS 2026 Evaluations and Datasets.
  • 2026-09 Starting my PhD at the University of Chicago!
  • 2026-01 PSALM-V: Automating Symbolic Planning in Interactive Visual Environments with Large Language Models accepted to ICRA 2026.
  • 2025-12 Graduated from USC with a B.S. in Computer Science!

Research (* indicates equal contribution)
Dual Exam

Dual Exam: Controlled Evaluation of Multimodal Agents and World Models


XGEN LabsXGEN Labs (83 authors, including Miaosen Chai)
Preprint
arXiv / Website / BibTeX

A paired benchmark that evaluates multimodal agents under fixed environment responses and world models under fixed action interventions across physical, digital, social, and scientific domains, revealing uneven task performance and failures to preserve task-relevant state.
PreciseCoder

PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards


Yunkai Zhan, Miaosen Chai, Shangshang Wang, Jike Zhong, Yuqing Yang, Deqing Fu, Serina Chang, Wang Bill Zhu
Preprint
arXiv / Code / Model / Website / BibTeX

PreciseCoder trains debugging models with RL using an edit-level F1 reward alongside unit tests, yielding faithful minimal fixes instead of over-editing, while preserving pass rate and generalizing beyond the training distribution.
Precise Debugging Benchmark

Precise Debugging Benchmark: Is Your Model Debugging or Regenerating?


Miaosen Chai*, Wang Bill Zhu*, Shangshang Wang, Yejia Liu, Song Bian, Honghua Dong, Willie Neiswanger, Robin Jia
NeurIPS 2026 Evaluations and Datasets
arXiv / Code / Website / BibTeX

A plug-and-paly data engine and benchmark for distinguishing whether a model truly debugs an existing program or simply regenerates it, disentangling repair capability from code generation with new edit-level metrics.
PSALM-V

PSALM-V: Automating Symbolic Planning in Interactive Visual Environments with Large Language Models


Wang Bill Zhu, Miaosen Chai, Ishika Singh, Robin Jia, Jesse Thomason
ICRA 2026
arXiv / Website / BibTeX

PSALM-V automates the induction of symbolic planning models directly from interactive visual environments, letting LLM-driven agents plan reliably without hand-crafted symbolic language for robotics.
Scientific Hypothesis Generation with Mamba

Exploring Scientific Hypothesis Generation with Mamba


Miaosen Chai*, Emily Herron*, Erick Cervantes, Tirthankar Ghosal
NLP4Science @ EMNLP 2024
Anthology / Code / BibTeX

An early exploration of state-space models for open-ended scientific hypothesis generation, examining how Mamba compares to transformer baselines on grounded scientific text.