Amazon-Virginia Tech initiative names Amazon Fellow and commits support to five faculty-led projects
Entering its fifth year, the Amazon-Virginia Tech Initiative for Efficient and Robust Machine Learning has chosen Hani Alomari, a Ph.D. student in computer science, as an Amazon Fellow for the 2026-27 academic year and has awarded support to five research projects led by Virginia Tech faculty.
Launched to advance research and innovation in artificial intelligence and machine learning, the initiative is funded by Amazon, housed in the College of Engineering, and directed by researchers at the Sanghani Center for Artificial Intelligence and Data Analytics.
All awardees are selected by an advisory committee of Virginia Tech faculty and Amazon researchers.
Amazon Fellowship
The Amazon Fellowship is awarded to Virginia Tech doctoral students pursuing educational and research experiences in AI-related fields who are recognized for their scholarly achievements and potential for future accomplishments. In addition to receiving funding for their work, fellows have an opportunity to interview for an Amazon internship intended to provide them with a greater understanding of industry and use-inspired research.
Alomari’s research lies at the intersection of computer vision and natural language processing, with a focus on how vision-language models and multimodal large language models represent and reason about visual content that may have multiple valid interpretations. His work explores how multimodal models can preserve diverse perspectives rather than collapse them into a single representation, with applications to cross-modal retrieval, visual reasoning and grounding, and image evaluation.
He is currently interested in extending these ideas toward personalized multimodal that can reason and iteratively improve retrieval, evaluation, and recommendations based on the user's intent. Advised by Chris Thomas, Alomari is particularly focused on multi-agent systems that can maintain distinct interpretations and adapt their behavior to the perspective, intent, and preferences of the active user.
Faculty-led Projects
The initiative also supports faculty work that is revolutionizing the way the world uses and understands artificial intelligence and machine learning technology.
The awardees and their projects are:
Ruoxi Jia, assistant professor in the Bradley Department of Electrical and Computer Engineering and core faculty at the Sanghani Center, received funding for “SpecRx: Specification-Grounded Trace Diagnosis for Self-Improving Agents.” AI agents are moving from answering isolated questions to completing long-horizon tasks through planning, tool use, retrieval, memory, code execution, and interaction with external environments. This shift creates a debugging problem: when an agent fails, the final error is often only a down-stream symptom. The decisive mistake may occur much earlier, when the agent misunderstands the task, skips a precondition, uses a tool in the wrong context, trusts stale evidence, or commits to a path that later becomes impossible to recover from. This project will deliver a focused benchmark for specification-grounded diagnosis of failed tool-using agents; a SpecRx diagnostic pipeline that identifies violated requirements and first progress-losing steps; a comparison against raw-trace and generic self-reflection baselines; and an Amazon-facing prototype API/dashboard for trace triage and repair scoping. The success criterion is a concrete diagnostic artifact that can reduce manual trace inspection and make self-improvement loops more data-efficient.
Murat Kantarcioglu, professor in the Department of Computer Science and core faculty at the Sanghani Center, received funding for “Toward Robust, Reliable, and Secure Evaluation of Agentic AI Systems.” Agentic AI systems which plan, invoke external tools, maintain persistent memory, and coordinate with one another to complete multi-step tasks, are being deployed at scale before the field has developed evaluation infrastructure adequate for this shift. Existing benchmarks are often unreliable, evaluation metrics can be unstable, and the security properties that determine whether these systems are safe to operate still lack a systematic and rigorous testing framework. This project addresses these gaps through a unified four-thrust research program: Contamination-Resistant Adaptive Benchmarks; Domain-Informed, Multi-Dimensional Evaluation; Reliability-Aware Evaluation Protocols; and Agentic Security and Testing for Reliable AI (ASTRA). The project will deliver an open-source adaptive benchmarking suite, a security-testing package, domain-specific evaluation feature libraries, and standardized protocols for reliability assessment and reporting.
Gretchen L. Matthews, professor in the Department of Mathematics and director of the Southwest Node, Commonwealth Cyber Initiative, with co-principal investigators Delio Jaramillo Velez, former postdoctoral associate, Hiram H. López, associate professor, and Kirsten Morris, postdoctoral researcher, all in the Department of Mathematics; and Ruoxi Jia, assistant professor in the Bradley Department of Electrical and Computer Engineering and core faculty at the Sanghani Center, received funding for “Toward Integrated Federated In context Learning (Fed ICL) and Contribution Evaluation for Responsible Generative AI.” Adoption of generative AI is moving rapidly from simple chatbots toward autonomous AI agents capable of reasoning, using tools, accessing enterprise data, and executing business workflows. This creates major opportunities for technology companies, banks, insurers, healthcare organizations, and customer service platforms to scale automation, reduce operational costs, and improve responsiveness. At the same time, it introduces new challenges related to reliability, safety, governance, and trust, particularly in regulated or high-risk environments where failures can have significant operational and reputational consequences. This project aims to develop responsible, contribution-aware Fed ICL mechanisms that improve enterprise AI agents while balancing utility, safety, fairness, and privacy. By integrating federated learning principles, in-context adaptation, agent orchestration, guardrail enforcement, and privacy-preserving contribution evaluation, this work seeks to provide a robust foundation for collaborative, enterprise-grade AI systems.
Dimitrios S. Nikolopoulos, John W. Hancock Professor in the Department of Computer Science, and Harpreet S. Dhillon, W. Martin Johnson Professor of Engineering in the Bradley Department of Electrical and Computer Engineering, and associate dean for Research and Innovation in the College of Engineering, received funding for “Eliminating Redundant Computation in Multi-Agent AI Systems.” Hard problems are increasingly handled not by a single AI model but by a team of them -- a planner, assistants that search and read, and a reviewer that checks the answer. But these agents cannot share what they have already worked out, so each reads the same material from scratch and repeats work a teammate finished seconds earlier. That waste is now the biggest single cost of running AI at scale, and it grows as the teams grow. The project is building software that lets these models hand their working memory directly to one another, even when their tasks are only similar rather than identical and even when the models differ in size. Because reusing the wrong information can quietly spoil an answer, the system learns, message by message, to judge when sharing is safe. Early results suggest about 40 percent of the wasted computation can be cut, with less cost, less energy, and faster, higher-quality answers.
Tu Vu, assistant professor in the Department of Computer Science and core faculty at the Sanghani Center, received funding for “Efficient and Scalable Agent Intelligence through Modular and Evolving Skills.” Agentic systems are typically built on foundation models that largely encode capabilities in their parameters and rely on task-specific information at inference time to perform specialized tasks. While this paradigm has produced many state-of-the-art results, agents must repeatedly rediscover solutions rather than accumulate knowledge and capabilities across tasks and domains. Recent work suggests that skills, which are structured packages of procedural knowledge such as instructions, scripts, and resources, offer an alternative path forward. This approach is appealing, since agents can discover task-specific expertise and encode it into reusable skills that they can load on demand to improve task performance. The goal of this project is to develop a skill-centric learning framework that facilitates the acquisition, composition, transfer, and evolution of modular skills, thereby enabling agents to reliably build and refine capabilities over time.
“The use of AI in our everyday life has grown significantly since we launched the initiative in 2022,” said Naren Ramakrishnan, University Distinguished Professor, director of the Sanghani Center, and director of the initiative.“This year’s awardees’ projects focus on efficiency, security, and trustworthiness of agentic workflows.”
"The Amazon-Virginia Tech Initiative is a model for how industry and academia can push the boundaries of AI together. Supporting five outstanding faculty-led projects and naming a new Amazon Fellow underscores our belief that breakthrough innovation comes from close, sustained partnership. We're proud to help accelerate research that is reshaping how the world understands and applies machine learning,” said Anand Rathi, Director of Software Development, AWS SageMaker AI.