Xiaodong (Will) An

Xiaodong An

Hi, I am a Ph.D. student in Physics at the Georgia Institute of Technology, advised by Prof. Flavio H. Fenton in the Chaos Lab, where I am also finishing an M.S. in Computational Science and Engineering (both expected Dec 2027). Before Georgia Tech I received my B.S. in Physics from The Hong Kong Polytechnic University in 2022, where I worked with Prof. Chi-Hang Lam on glassy dynamics.

My research interests include evaluation and LLM-as-a-judge, agentic systems and tool use, and AI for science.

I am currently building SciAgent Bench, a benchmark for agentic AI in scientific discovery with Scale AI and Georgia Tech, and working on RLVR post-training for PDE solver code synthesis. In summer 2025 I was a machine learning engineer at Scam AI in Berkeley, on adversarial data generation and a mixture-of-experts face-swap detector.

Contact: xan37@gatech.edu
Links: CV|GitHub|LinkedIn

publications

WEB-PDE-LLM agent pipeline

WEB-PDE-LLM: A Multi-Agent Framework for Automated Real-Time Web-Based GPU PDE Solving

Xiaodong An, Guangze Luo, Flavio H. Fenton

Under review, NeurIPS 2026 Workshop on Sim2Science

Agents turn a plain text PDE description into a WebGL solver that runs on the GPU in the browser. They check the CFL condition before running, then repair the solver from its own error output.

LLM judge scores per task

Do LLM Judges Evaluate What We Ask Them to Measure? A Controlled Benchmark of Complexity and Chaos in Scientific Time Series

Afrouz Delshad*, Xiaodong An*, Elizabeth M. Cherry, Flavio H. Fenton  (* equal contribution)

Under review, NeurIPS 2026 JUDGe Workshop

We build time series whose complexity and chaoticity are known by construction, then ask LLM judges to score them. The judges agree with each other more than with the quantity they were asked to measure.

PolicyAgent framework and PHPC-Bench

An Agentic Framework for Auditable Public Health Policy Communication

Yansong Wen, Chengyun Fan, Xiaodong An, et al.

Under review, Nature Health

Epidemiologist, clinician, and manager agents draft a public health message from surveillance and projection data. PHPC-Bench then checks the writing, its agreement with CDC and WHO guidance, and whether every number cited matches the source data.

Cosmetic attacks on age estimation

Can a Teenager Fool an AI? Evaluating Low-Cost Cosmetic Attacks on Age Estimation Systems

Xingyu Shen, Xiaodong An, et al.

arXiv:2602.19539, 2026

A beard, grey hair, makeup, or wrinkles added in an ordinary photo editor, with no access to the model, push face based age estimators over the adult threshold for minors. We measure the age shift per attack on 329 subjects.

Spiral wave dynamics, panels A to F

Quantifying the Complex Spatiotemporal Chaos of Cardiac Fibrillation in Ionic Models Across Parameter Regimes

Xiaodong An, et al.

arXiv:2508.14303, 2025

Temporal and spatiotemporal Lyapunov exponents for spiral wave dynamics in cardiac ionic models, from circular cores through meander to hypermeander, as excitability and sodium conductance are varied.

Swap initiators in a glassy particle system

Diffusion-Coefficient Power Laws and Defect-Driven Glassy Dynamics in Swap Acceleration

Gautham Gopinath, Chun-Shing Lee, Xin-Yuan Gao, Xiao-Dong An, Chor-Hoi Chan, Cho-Tung Yip, Hai-Yao Deng, Chi-Hang Lam

Physical Review Letters 129, 168002 (2022)

Only a fraction of the particles is allowed to swap. The diffusion coefficients then follow power laws in that fraction, with exponents that depend on temperature, and the regular particles show ordinary glassy behaviour driven by a sparse set of swapping defects.