Menghan Xu

mx253@cornell.edu

my_pic_new.jpg

Feel free to reach out!

I’m a master’s student at Cornell University, where I do research in machine learning. I’m particularly interested in large language models, world models, and reinforcement learning. I received my B.S. in Statistics from the University of Science and Technology of China.

Beyond research, I enjoy algorithms and topics like LP duality, the hedge algorithm, and game theory. I also love learning about astronomy and astrophysics. In my free time, I’m usually listening to indie music, singing, playing guitar, or watching a movie, and I’m always up for a board game with friends.

I’m currently working with Prof. Kevin Ellis on a super cool project exploring temporal abstraction in world models.

publications / preprints

  1. COLM ’26
    fig_multihop_overview.png
    Co-Evolving Structured Knowledge and Reasoning in Language Models
    Ryan Thomas Noonan*, Linxi Zhao*, Menghan Xu*, Akanksha Sarkar, Mihir Mishra, Dongyoung Go, Kilian Q. Weinberger, Yoav Artzi, and Jennifer J. Sun
    * Equal contribution
    In Conference on Language Modeling (COLM), 2026
    arXiv Code Website
  2. In Submission
    Spatial Induction Heads: In-Context Learning of Multidimensional Cellular Automata
    Kimia Kazemian*, Menghan Xu*, John Thickstun, and Sarah Dean
    * Equal contribution
    2026

posts

Dec 18, 2025 Submodular Functions: Theory, Exact Minimization, and Approximate Maximization
Nov 4, 2025 Presentation: Decision Transformer: RL via Sequence Modeling

projects

Uncertainty-Aware Collision Avoidance through Safe Reinforcement Learning
Uncertainty-Aware Collision Avoidance through Safe Reinforcement Learning
Proposed a constrained MDP framework using lagrange PPO and behavior cloning for navigating unsignalized intersections under uncertainty. The method achieves a 100% success rate and zero-shot generalization to out-of-distribution agent behaviors (e.g., zigzagging), validated via high-fidelity simulation and hardware-in-the-loop experiments on Jackal robots.
My Smart Memory
My Smart Memory
Developed a full-stack RAG application using Azure SQL database and OpenAI. Features vector similarity search for long-term conversation memory, implemented with React, Node.js, and LangChain to enable context-aware AI interactions.