Huaicheng Li and Chris Thomas, assistant professors in the Virginia Tech Department of Computer Science, have been named Google Research Scholars, a highly competitive international honor that supports early career professors conducting world-class research.

Thomas is also core faculty at the Sanghani Center for Artificial Intelligence and Data Analytics.

The Research Scholar Program awards unrestricted gifts in 12 categories to institutions around the world to fund promising research across key areas in computer science and engineering.

Faster, smarter operating systems for data centers

Li’s award from the program’s systems category supports his work on operating systems.

“My focus is on building systems that harness the power of emerging storage and memory hardware to improve the performance and efficiency of data-intensive applications by rethinking how data is managed at the system level and exploring innovative ways to manage memory in data centers that use tiered memory architectures,” Li said.

Typically these data centers include two types of memory: a performance tier that is fast but limited in capacity and a capacity tier that offers increased storage but operates more slowly. Traditional operating systems try to boost performance by placing frequently accessed “hot” data in the fast memory and “cold" data in the slower memory, but recent research showed that approach is not always effective.

“Just because data is accessed often doesn't mean it has a major impact on performance,” Li said.

Real-world impact

Google’s support will help Li develop a lightweight, accurate method to predict the performance benefit of storing a piece of data in fast memory, an approach that could reshape how operating systems allocate memory in data centers. 

The predictor will allow the researchers to:

  • Redesign how operating systems allocate memory and move data between tiers, prioritizing the most performance-critical data for fast memory
  • Improve responsiveness while reducing operational costs for a wide range of systems and applications, including machine learning and artificial intelligence, enterprise platforms, and high-performance computing environments
  • Help researchers and developers working with next-generation memory technologies gain new tools and insights to build faster, more efficient computing systems

Li's Ph.D. students Jinshu Liu, Hamid Hadian, and Hanchen Xu are supporting the project. The team’s previous work includes a large-scale study of 265 real-world workloads across a variety of server and memory configurations, presented at the 2025 Association for Computing Machinery International Conference on Architectural Support for Programming Languages and Operating Systems. Li said the study highlighted the performance issues with slow memory and inspired the current research direction.

“This project is an exciting opportunity to challenge long-standing assumptions in operating system design, especially as memory technologies continue to evolve rapidly,” Li said. “We hope our work will help pave the way for more intelligent and adaptable computing systems in the years ahead.”

Making digital assistants safer

With research at the intersection of computer vision, natural language processing, and multimedia, Thomas is striving to build artifical intelligence (AI) systems that are not only capable of complex multimodal reasoning at scale, but also safe, robust, and trustworthy when deployed in real-world settings. 

Thomas will serve as principal investigator for a project selected in Google’s privacy, safety, and security category to research a new class of artificial intelligence systems called agentic AI.

“These AI systems go beyond answering questions," Thomas said. "They can take actions using tools, such as an internet browser, app, etc., on the user’s behalf based on spoken natural language instructions.”

But that comes with risk. For example, you could tell your AI assistant, “Help me dispute this medical bill,” and it would go online, log into your insurance portal, pull up your claim history, draft a message to customer support, and submit the form automatically.

“This long sequence of actions and reliance on memory make them uniquely vulnerable to subtle forms of manipulation called adversarial attacks in AI research," Thomas said. "A malicious site could, for example, insert tiny alterations in images on the page that are imperceptible to the naked eye and steer the agent off course.

Guarding against adversarial attacks

To make AI automation more accessible and trustworthy, Thomas and his team, including graduate students Alvi Ishmam, Aafiya Hussain, Zaber Hakim, and Najibul Sarker, have the following goals:

  • Identify and stress-test vulnerabilities, developing new types of adversarial attacks that simulate how a bad actor might manipulate the online environment to confuse or control an AI agent
  • Analyze the risks of these attacks to develop a clearer understanding of their impact
  • Address safety issues so that agentic AI systems see broader use
  • Guide future research toward the most important security challenges

As part of the Google Research Scholar program, the project will provide student researchers the opportunity to collaborate with other groups across the country and will feature meetings with highly influential researchers in the security and AI communities at Google.

Agentic AI, Thomas said, has the potential to benefit a broad range of users because it requires no technical expertise to use. People interact through natural language, turning what used to be multistep, manual tasks into automated actions. 

“This kind of accessibility could dramatically expand who benefits from automation,” he said. “But for agentic AI to succeed at scale, users need to trust it, so that is why our research supports not just end users, but anyone building or deploying these systems in real-world applications.”

Contact:

Tags

Share this story