ziruiz [AT] comp.nus.edu.sg
ryan_zzr [AT] outlook.com
Zirui Zhao is a Ph.D. student in Computer Science at the National University of Singapore, advised by Prof. Wee Sun Lee. He also works closely with Prof. David Hsu. He is doing research in decision-making under uncertainty, robotics, and their intersection with Machine Learning, NLP, and Large Language Models. He received his Bachelor's Degree at Xi'an Jiaotong University. He used to work with Prof. Pengju Ren during his undergrad study and Prof. Ding Zhao when visiting CMU.
Google Scholar  /
I prefer to think deeply and dig into the simple underlying essence of a complex problem/method. I am interested in solving large-scale decision-making problems in complex environments and their application in robotics/embodied AI. I am also interested in leveraging search algorithms/external tools/structured memory to boost LLM reasoning capabilities, especially in planning/NLP problems.
On the Effective Complexity of LLM for planning and reasoning
Wee Sun Lee (*Equal contribution, listed in alphabetical order)
We analyse the empirical behaviours of LLMs in reasoning and planning problems from the prepectives of sample and computational complexity.
Large Language Models as Commonsense Knowledge for Large-Scale Task Planning
Wee Sun Lee,
Also in RSS LTAMP Workshop, Best Paper Runner-up, 2023
project page /
We use Large Language Models as both the commonsense world model and the heuristic policy within Monte Carlo Tree Search. LLM's world model provides with MCTS a commonsense prior belief of states for reasoned decision-making. The LLM's heuristic policy guides the search to relevant parts of the tree, substantially reducing the search complexity.
Differentiable Parsing and Visual Grounding of Natural Language Instructions for Object Placement
Wee Sun Lee,
Also in CoRL LangRob Workshop, 2022
project page /
IEEE Xplore /
We proposed a novel method, ParaGon, for language-conditioned object placing. ParaGon integrates a parsing algorithm into an end-to-end trainable neural network. It is data-efficient and generalizable for learning compositional instructions, and robust to noisy, ambiguous language inputs.
Active Learning for Risk-sensitive Inverse Reinforcement Learning
Risk-sensitive inverse reinforcement learning provides an general model to capture how human assess the distribution of a stochastic outcome when the true distribution is unknown (ambiguous). This work enables an RS-IRL learner to actively query expert demonstrations for faster risk envelope approximation.
Visual Semantic SLAM
Zirui Zhao, Yijun Mao, Yan Ding,
Pengju Ren, Nanning Zheng
Semantic SLAM projecting semantic meaning into 3D point clouds generated by ORB SLAM algorithm.
I am particularly interested in teaching, especially in the field of AI. I have been a teaching assistant for CS1010E (Programming Methodology), CS3245 (Information Retrieval), and teaching tutorials of CS3263 (Foundations of Artificial Intelligence) for two consecutive semesters. I was selected for the Teaching Fellowship Award in AY23/24-Sem1.