CV

Curriculum Vitae

Xudong Zhu

I am a PhD student in Computer Science at The Ohio State University, advised by Prof. Zhihui Zhu. My research examines how large language models represent concepts and how these representations evolve during generation and training. I am particularly interested in the linear structure of these representations and its role in understanding and steering model behavior. I also study how different training methods, such as SFT and RL, shape this structure. I received my B.S. in Computer Science from UESTC, where I worked with Prof. Zhao Kang on relation extraction and Prof. Dong Hao on graph learning.

Download CV zhu.3944@osu.edu xudongzhu.com Columbus, Ohio

Academic path

Education

  1. Present

    Ph.D. in Computer Science and Engineering

    The Ohio State University

    Columbus, OH, USA

    Department of Computer Science and Engineering. Advised by Prof. Zhihui Zhu.

  2. B.S. in Computer Science

    University of Electronic Science and Technology of China

    Chengdu, Sichuan, China

    School of Computer Science and Engineering.

    GPA 3.98 / 4.00

Research

Research Experience

Doctoral Research · The Ohio State University

Advisor: Prof. Zhihui Zhu

  • Concept representations. Study how LLMs represent concepts, focusing on their linear structure, bidirectional features, and connections to model interpretability and behavioral control. Related work: AbsTopK (ICLR 2026).
  • Generation dynamics. Study how concept representations evolve during generation and how steering propagates their influence on subsequent model behavior. Related work: Self-reflection (TMLR 2026); Prefix Steering (arXiv 2026).
  • Training dynamics. Study how training methods such as SFT and RL reshape concept representations, their linear structure, and their influence on behavior over the course of training. Related work: Post-training Dynamics (ongoing).

Research

Publications

  2. AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features

    ICLR 2026

    Xudong Zhu, Mohammad Mahdi Khalili, Zhihui Zhu

    Sparse autoencoders (SAEs) are widely used for LLM interpretability, but existing variants often impose non-negativity that prevents single features from representing bidirectional concepts. We derive SAE variants from unrolled proximal gradient updates, identify this structural limitation, and propose AbsTopK SAE with magnitude-based hard thresholding. Across four LLMs and seven probing/steering tasks, AbsTopK improves reconstruction and interpretability while enabling single features to encode contrasting concepts.

  3. From Emergence to Control: Probing and Modulating Self-Reflection in Language Models

    TMLR 2026

    Xudong Zhu, Jiachen Jiang, Mohammad Mahdi Khalili, Zhihui Zhu

    We study the emergence and control of self-reflection in large language models. Our probing method reveals that pretrained models already contain a latent capacity for reflection, which can be amplified without additional training. By identifying and manipulating a “self-reflection vector” in activation space, we achieve bidirectional control over reflective behavior, improving reasoning accuracy or reducing computation as needed. This work deepens understanding of self-reflection and demonstrates how model internals can enable precise behavioral modulation.

  4. FCDS: Fusing Constituency and Dependency Syntax into Document-Level Relation Extraction

    LREC-COLING 2024

    Xudong Zhu, Zhao Kang, Bei Hui

    We introduce FCDS, a document-level relation extraction model that fuses constituency and dependency syntax. By combining sentence-level aggregation from constituency trees with dependency-based graph reasoning, FCDS better captures cross-sentence relations between entities. Experiments across multiple domains show significant performance gains, highlighting the effectiveness of integrating both syntactic views.

  5. Alleviating subgraph-induced oversmoothing in link prediction via coarse graining

    Neurocomputing, 2025

    Xudong Zhu, Dong Hao, Ziqin Gao, Liming Pan

    We address the oversmoothing problem in link prediction caused by repetitive high-degree nodes across subgraphs. Our method introduces a coarse-graining strategy that merges strongly correlated nodes, yielding more diverse receptive fields and reducing subgraph size. This not only mitigates oversmoothing but also improves scalability and efficiency of GNN-based link prediction.

Teaching

Teaching Experience

Teaching Assistant · The Ohio State University

CSE 3521: Introduction to Artificial Intelligence · Autumn 2025

Service

Academic Service

Reviewer: NeurIPS, ICLR, ICML · 2026

Recognition

Honors & Research Support

Selected software

Open Source

TypeScript

LLMUsage

Local-first LLM subscription usage for Hyprland, OpenAI, GLM, Grok, and Claude in your bar.

  • CLI
  • Quickshell

Lua

dotfiles

Personal dotfiles for fish, Neovim, and OpenCode, managed with GNU Stow.

  • Neovim
  • OpenCode