Princeton University · Electrical and Computer Engineering

Suozhi Huang

Ph.D. Student & Researcher

I am a Ph.D. student in Electrical and Computer Engineering at Princeton University, advised by Prof. Mengdi Wang. My research focuses on building and evaluating LLM agents for expert-level reasoning and decision-making in complex, dynamic environments. I am particularly interested in agentic reasoning, model adaptation, and reliable evaluation in real-world and scientific domains. I am also interested in Prof. Wang’s LabOS initiative and its vision of AI-XR co-scientists that collaborate with human researchers in real-world scientific workflows.

My recent work spans agent evaluation and future prediction, multimodal agent systems, LLM architectures, and formal theorem proving. Before joining Princeton, I received my B.E. in Computer Science from Tsinghua University’s Yao Class. I began my research in collaborative robotic perception and later worked on AI for science and formal mathematical reasoning.

I have been fortunate to work with Prof. Anima Anandkumar at Caltech; Dr. Zheng Yuan and Prof. Dahua Lin through InternLM-Math; and Prof. Chen Feng and Dr. Yiming Li at NYU. At Princeton, I have also worked with Prof. Dmitry Abanin on quantum error correction and cooling.

Research interests

From learning and reasoning to agents that act in the world.

Agents, Memory &
Self-Improvement

I build research-agent harnesses, persistent memory, and evaluation frameworks that connect experimentation with feedback. My work also examines how to measure agent capabilities on expert-level financial and forecasting tasks.

LLM Learning,
Reasoning & Inference

I study how language models learn, search, and use computation. In multi-teacher on-policy distillation, I investigate optimizer-dependent gradient dynamics and weight updates. In LoopFill, I trained DFlash drafters, explored smaller and faster draft models, and reproduced the LoopAhead baseline.

My earlier work includes critic-guided search and expert-iteration data construction for InternLM2.5-StepProver, and proof-progress prediction and LeanCopilot integration for LeanProgress. I am also a coauthor of Gated Attention, a NeurIPS 2025 Best Paper.

World Models &
AI for Science

I work on digital twins and 4D reconstruction of scientific laboratories, including trajectory construction for robot simulation training data. My broader work spans collaborative perception, weather forecasting, molecular properties, and quantum systems.

Publications & projects

Google Scholar

2026

LabWorld: Learning Interactive 4D Worlds for Scientific Experimentation

Interactive 4D worlds for scientific experimentation.

Suozhi Huang*, Kelsey Fu*, Binxu Li, Luna Lyu, Kaize Wu, Wenbo Lu, David Smerkous, Jiahan Li, Siqi Zhu, Zhizhou Sha, Jiacheng Guo, Zhixuan Liang, Shilong Liu, Le Cong, and Mengdi Wang

arXiv preprint · * Equal contributionPaperProjectCode
My contribution

Developed laboratory digital twins and 4D scene reconstructions, including trajectories for robot simulation training data.

2026

AQuA: Recursively Self-Improving Quantitative Trading Research Agents

Research agents that improve through experimentation, evaluation, and persistent memory.

Jiacheng Guo, Suozhi Huang, Yunlong Gao, Zihao Li, Jason Ge, Xu Kuang, and Mengdi Wang

arXiv preprintPaper
My contribution

Built the self-improving agent harness with persistent memory, alongside the financial data and evaluation framework.

2025

LeanProgress: Guiding Search for Neural Theorem Proving via Proof Progress Prediction

Guiding theorem-proving search with predictions of proof progress.

Suozhi Huang, Peiyang Song, Robert Joseph George, and Anima Anandkumar

Transactions on Machine Learning ResearchPaper

Author order shown follows the original arXiv submission.

My contribution

Generated proof-search tree data, trained proof-progress predictors, and integrated progress guidance into LeanCopilot.

2024

AI4Math 2025 Workshop · Best Paper Honorable Mention

InternLM2.5-StepProver: Advancing Automated Theorem Proving via Expert Iteration on Large-Scale LEAN Problems

Expert iteration and critic-guided search for formal mathematical reasoning.

Zijian Wu*, Suozhi Huang*, Zhejian Zhou, Huaiyuan Ying, Jiayu Wang, Dahua Lin, and Kai Chen

Technical reportPaperCodeDataset
My contribution

Built the critic model, integrated critic-guided tactic selection, and constructed proof data through expert-iteration search.

* Equal contribution. Preprints are labeled separately from published work.