Haozhu Wang 

I'm an AI researcher working on Alignment and Reinforcement Learning for Llama models at Meta GenAI. I earned my Ph.D. degree in Electrical and Computer Engineering (Machine Learning track) from the University of Michigan, Ann Arbor. My research focuses on LLMs, reinforcement learning, alignment, agent, and AI for science. I'm open to collaboration with aspiring researchers and I frequently mentor students on AI research projects. If you find shared research interests and would like to discuss collaboration opportunities, please feel free to contact me via email!

Email  /  LinkedIn /  Google Scholar  /  GitHub

[Nov-2024] We presented our work LaRS: Latent Reasoning Skills for Chain-of-Thought Reasoning at EMNLP 2024.

[July-2024] Our work OptoGPT, a foundation model for optical inverse design, has been published as a cover article in Opto-Electronic Advances (IF: 14.1). The work has been reported in over 15 news outlets [News1] [News2] [News3].

[Dec-2023] Our work Graph Neural Prompting with Large Language Models is accepted by AAAI-24.

[Dec-2023] 4 workshop papers accepted by NeurIPS'23.

[Aug-2022] Our paper Dynamic prediction of work status for workers with occupational injuries: assessing the value of longitudinal observations has been published on Journal of the American Medical Informatics Association!

[Aug-2022] Our AWS Machine Learning blog on real-time fraud detection is online.

[July-2022] I served as a session chair in ICML 2022.

[Mar-2022] I joined AWS as a Research Scientist!

We developed an unsupervised method for discovering latent skills to guide the demonstration selection for in-context learning with large langauge models.

A knowledge graph prompting method for large language models to improve their commonsense and biomedical reasoning performance.

A comprehensive review of reinforcement learning applied to NLP and its healthcare applications.

We developed OptoGPT, the first foundation model for optical thin film structure inverse design. After being trained on a large dataset of 10 million optical thin film designs, OptoGPT demonstrates remarkable capabilities including: 1) autonomous global design exploration, 2) efficient designs for various tasks, 3) the ability to output diverse designs, and 4) seamless integration of user-defined constraints. We believe OptoGPT is a major leap towards accelerating optical science with foundation models.

We developed an hierarchical prompting method for transformer-based reinforcement learning models to enable efficient few-shot policy adaptation.

Using reinforcement learning, we designed and fabricated two multilayer thin film structures that can mimic the visual appearance of decorative chrome plating, serving as a environmentally friendly and multi-functional replacement.

We developed a forecasting model to predict return-to-work after occupational injuries based on longitudinal claim data. The model may allow case managers to better allocate medical resources and help speed up patients' recover process.

We propose a hybrid machine learning and optimization method that combines mixture density networks and particle swarm optimization for accurate and efficient structural color inverse design.

We provide extensive benchmarking results on accuracy, diversity, robustness for commonly used deep learning models in nanophotonic inverse designs. The findings can help researchers select models that best suit their design problems.

Training a novel sequence generation network with Proximal Policy Optimization for automatically discovering near-optimal optical designs.

Expert-yielded-estimates regularizer for incorporating expert knowledge into linear models.

Group-ordered-weighted lasso (GrOWL) for deep model compression.

The source code of this website is from Jon Barron.

(last update: 2024)