Xi Xiao

Hi👋 I'm Xi Xiao, a third-year Ph.D. student in Computer Science at the University of Alabama at Birmingham, cd-advised by Prof. Tianyang Wang and Prof. Min Xu from Carnegie Mellon University. I'm also a student researcher at Oak Ridge National Laboratory working with Dr. Xiao Wang since May 2025.

I am a core contributor to ORBIT-2, an exascale climate foundation model trained on the Frontier supercomputer. The work received the Best Paper Award at SC 2025 and is a ACM Gordon Bell Prize finalist. ORBIT-2 has since been integrated into NVIDIA’s enterprise climate AI stack, powering large-scale forecasting services that reach millions of users worldwide.

I am currently seeking a Summer 2026 internship focused on LLMs / MLLMs post-training stage research and applications, please feel free to connect me if my experience match your team:)

Email  /  CV  /  Scholar  /  GitHub  /  LinkedIn

News

Research

I have broad interests in computer vision, and language models. My recent work focuses on the post-training stage of large-scale models (LLMs/MLLMs/LVMs), including parameter-efficient fine-tuning (PEFT), reinforcement learning–based alignment, and model quantization. I aim to build efficient and robust intelligent systems that can perform reliably in extreme scenarios with limited data, limited compute, and limited storage resources. Some representative papers are highlighted.

An exascale vision transformer for high-resolution climate downscaling on Frontier, enabling accurate and efficient prediction of regional weather and climate extremes.

A comprehensive taxonomy and survey of visual prompt tuning and prompting for large vision models, covering learnable, generative, and non-learnable prompts across diverse tasks.

Instance-aware visual prompts that adapt to each image, mitigating overfitting and improving transferability of ViT-based classifiers under distribution shift.

A hybrid preference optimization framework that jointly preserves identity and motion dynamics for personalized text-to-video generation.

A correlation-aware latent space that jointly improves disentanglement and fairness in generative models through causal regularization.

A routed mixture-of-experts architecture for diffusion-based fMRI-to-image reconstruction, achieving strong cross-subject generalization and interpretable brain–model alignment.

An efficient fine-tuning method that dynamically allocates ranks to weight matrices based on both their global and local sensitivities. It leverages the second-order derivatives (Hessian Matrix) of the loss function to effectively capture weight sensitivity, enabling optimal rank allocation with minimal computational overhead.

A novel representation engineering approach that replaces explicit token-level demonstrations with a set of learnable Multimodal In-context Vectors directly injected into the residual streams of LVLMs.

Professional Experience

Research Intern, Oak Ridge National Laboratory (ORNL) 05/2025 – Present Knoxville, Tennessee, USA Working with Dr. Xiao Wang on large-scale climate models (e.g., ORBIT-2), efficient post-training, and exascale distributed training.

Ph.D. Student, University of Alabama at Birmingham (UAB) 01/2024 – Present Birmingham, Alabama, USA Advised by Prof. Tianyang Wang; research in efficient adaptation of large-scale models.

Awards & Honors

Miscellanea

Website adapted from Jon Barron's template.