Zonghuan Xu

Zonghuan Xu

Fudan University, School of Mathematical Sciences. B.S. in Mathematics and Applied Mathematics, Xianghui Plan, expected 2028. Exchange student at the University of Toronto, St. George, Fall 2026.

Contact: 2430XH10002@m.fudan.edu.cn

My research interests are machine learning theory, continual learning, trustworthy AI, and embodied AI.

My current work studies training-time scaling in gradual adaptation, forgetting under task distributions, VLA safety, and human-grounded LLM evaluation.

Zonghuan Xu portrait
cv scholar orcid github arxiv openreview researchgate

publications

* Corresponding author(s).

Conference Papers

DropVLA: An Action-Level Backdoor Attack on Vision-Language-Action Models

Xu, Z., Li, J., Zhao, Y., Zheng, X.*, Ma, X.*, and Jiang, Y.-G.

Accepted at IROS 2026. arXiv:2510.10932.

pdf arXiv code

Preprints

From Order to Distribution: A Spectral Characterization of Forgetting in Continual Learning

Xu, Z. and Ma, X.*

Preprint, 2026. arXiv:2604.13460.

pdf arXiv

The Sequential Price of Continual Learning

Xu, Z. and Ma, X.*

Manuscript, 2026.

pdf

Optimal Training-Time Scaling in Gradual Adaptation

Xu, Z.* and Harish, K.

Preprint, 2026. arXiv:2608.04927.

pdf arXiv

Hypothesis Testing with Conditional Queries: Learnability and the Value of Interaction

Xu, Z.*

Preprint, 2026. arXiv:2608.06262.

pdf arXiv

When Direct Prediction Fails: Evidence from LLM-Based Misinformation Risk Evaluation

Xu, Z., Zheng, X., Wu, Y., and Ma, X.*

Preprint, 2026. arXiv:2604.06820.

pdf arXiv

AttackVLA: Benchmarking Adversarial and Backdoor Attacks on Vision-Language-Action Models

Li, J., Zhao, Y., Zheng, X., Xu, Z., Li, Y., Ma, X.*, and Jiang, Y.-G.*

Preprint, 2025. arXiv:2511.12149.

pdf arXiv

Position Papers

Human Model: The Missing Piece Toward Trustworthy AGI

Xu, Z., Ma, X.*, and Jiang, Y.-G.*

OpenReview Archive, 2026. Position paper.

pdf OpenReview

current questions

The questions below reflect some of my current interests. My previous work has given me experience in both theoretical and empirical research, and I am interested in exploring these questions from either perspective.

1. Understanding Humans as a Learnable Object

I am interested in whether AI can develop a general and transferable understanding of humans themselves.

An initial formulation of this question appears in my position paper.

2. Beyond Human Preference

I am interested in how AI systems should be evaluated when human judgment is no longer a sufficient universal measure of capability.

3. Learning Across Three Timescales

I am interested in learning across ordinary, continual, and meta-learning timescales, and in what these levels may reveal about the efficiency gap between human and machine learning.

4. Human–AI Co-evolution and Alignment

I am interested in bidirectional alignment between humans and AI, particularly as AI becomes part of human memory, reasoning, and decision-making. How can we design systems that support this co-evolution while treating identity continuity and human agency as alignment objectives?

I explore this direction in my essay Beyond Control: Becoming the Intelligence We Build.

long-term direction

In the longer term, I hope to learn how to connect empirical research, theory, and method design more closely, so that observations can inform theory and theoretical understanding can eventually lead back to better methods.

connect

essays