Xudong Shen

I obtained my Ph.D. in AI from National University of Singapore in 2024, where I was advised by Prof. Mohan Kankanhalli. I earned my bachelor's degree from Zhejiang University, China, in 2019. During 2022–2023, I interned at Sea AI Lab in Singapore, working with Chao Du and Tianyu Pang. In 2024–2025, I co-founded a venture-backed AI startup.

I am passionate about scaling Reinforcement Learning for multi-modal, long-horizon, complex real-world tasks. My earlier work focused on AI fairness, robustness, safety, and governance, where I took an evaluation-driven approach: stress-testing systems and translating findings into model improvements.

Email  /  CV  /  Scholar  /  Twitter  /  Github

Research

We developed a method to optimize diffusion models for any differentiable objective defined on the generated data, where score/noise prediction and RL fail. We applied it to enhance & control output diversity in text-to-image generation.

Image+text prompts trigger more safety failures than text-only.

Probes multilingual stereotypes and its cross-lingual transfer in LLMs.

Shows when larger models consistently perform worse; analyzes failure modes.

Large-scale eval that reveals where LLM capabilities scale well & where they don't.

Stress-tested LLM robustness using 100+ natural-language augmentations.

Instruction-tuning on 1.6K tasks boosts zero-shot unseen-task performance.

Learns representations with multiple approximate robustness guarantees that transfer to even unseen tasks. We applied it to debias face representations.

Learns representations with built-in interpretability (i.e., part–whole relationships) via capsule networks.

Time-to-event modeling to predict default & profitability on million-scale lending data.

Maps technical mechanisms that make AI easier to regulate in practice.

Analyzes how Generative AI & foundation models interact with platform regulation.

Feel free to steal this website's source code from Jon Barron.