Xuzhong Wang

Xuzhong Wang

Hi! I am an undergrad at Columbia University studying Computer Science. I am currently interested in machine learning research, specifically in post-training LLMs and AI safety. Before this, I did a brief stint of research in theoretical computer science, with a focus on the diameter of graph coloring reconfiguration problems.

  • Interests
  • Large Language Models
  • Agents
  • Reinforcement Learning
  • Graph Theory

Research

Broadly, I am interested in how agents can learn and adapt in complex environments. In parallel, I study and evaluate the representations in LLMs that emerge post-training—focusing on how models encode, interpret, and ground knowledge about the physical and conceptual world. Here are a few more concrete directions I am currently exploring:

  • Multi-Objective Alignment

    Develope techniques to improve how models navigate through distinct objectives — accuracy, cost, and format — in reasoning, coding, and tool usage.

  • Unlearning

    Develop techniques to erase the impact of certain data for ML models during training. Probe the LLM to see how "forget" acutally works in comparison to "confusion".

Selected Publications & Preprints

  • EMNLP 2026 Findings Multi-Objective RL

    Towards Reliable, Generalizable, and Specific In-Context Knowledge Editing via Multi-Objective Reinforcement Learning

    X. Wang, M. Jiang, T. Nair, G. Bhusal, Y. Zhang, H. Chen

    arXiv Code

The complete record is on Google Scholar.

View full list on Google Scholar

It always takes longer than you expect, even when you take into account Hofstadter's Law.

— Hofstadter's Law

Industry Experience

  • AnytimeAI

    ML engineer intern

    Built LLM agent pipelines for legal document analysis. An agent routing tool cut inference costs ~60%.

  • Bestlink Technology

    ML engineer intern

    Built an OpenCV data retrieval pipeline and a LangChain-based QA agent for consulting clients.

Projects

MO-IKE

Python · PyTorch

Official implementation of our multi-objective in-context knowledge editing work. An RL-trained retriever balances edit reliability, generality, and specificity against one another.

Repository Paper

Akita

Go · SQLite

An open-source computer architecture simulator I help maintain. I contributed the data recording package under Prof. Yifan Sun, for tracking simulation data across experiments.

Repository

Graph Recoloring

React · Vite

An interactive playground for graph coloring reconfiguration.

Repository Website