Zhiyi Shi

I am a first-year PhD student in Computer Science at University of Illinois Urbana-Champaign (UIUC), co-advised by Prof. Jimeng Sun and Prof. Quanzheng Li from Harvard MGH.

Prior to that, I received my Master's Degree from Carnegie Mellon University and my Bachelor's Degree from Southeast University.

My research interests include Agentic AI, RL, Parameter Efficient Fine-Tuning.

I am actively seeking Summer 2026 Internship!

Email: zhiyis3 [AT] illinois [DOT] edu

CV  /  Google Scholar  /  GitHub  /  LinkedIn

(*Equal Contribution)

This paper provides a comprehensive framework for Agentic AI adaptation, categorizing how systems evolve to handle complex tasks. It introduces a unified taxonomy that distinguishes between Agent Adaptation (refining the model's internal reasoning/planning) and Tool Adaptation (optimizing external components like APIs and memory). By analyzing trade-offs in cost, modularity, and generalization, the work offers a practical roadmap for building more reliable and autonomous AI agents.

We study task-specific directions (TSDs) as the key subspace that drives performance gains in parameter-efficient fine-tuning, and propose a unified framework to define, analyze, and utilize these directions in PEFT. Building on this insight, we introduce LoRA-Dash and LoRA-Init, and further combine them into LoRA-TSD, achieving consistent and significant improvements over standard LoRA across extensive experiments.

DualEdit edits both textual and visual modalities in vision-language models at their most sensitive layers. A gating mechanism in the textual modality enables knowledge updates without harming the model's original abilities.

We propose LoRA-Dash, a framework that leverages task-specific directions (TSDs) to maximize fine-tuning efficiency and improve task performance.

We propose a generalization of matrix-based PEFT to higher-dimensional parameter spaces using Lie group modeling, ensuring structure-preserving updates and demonstrating improved performance across vision and language tasks.

MoRA is a computationally efficient method for multi-modal disease diagnosis that adapts to missing modalities using modality-specific low-rank projections.

Website Template