Institution profile

Rochester Institute of Technology

Academic institutionnorthamerica · us
Official website
Research library391linked papers
Opportunities0open roles
Selected work

Representative Papers

The Unpaid Toll: Quantifying the Public Health Impact of AI

Dec 09, 2024arXiv.org

The environmental health impacts and associated environmental injustice of AI’s full lifecycle—from semiconductor manufacturing to datacenter operations—remain poorly quantified. Method: We develop the first integrated AI health impact quantification framework, combining multi-source emission inventories, life cycle assessment, atmospheric transport modeling, health risk assessment, and spatial environmental justice statistics. Contribution/Results: Training Llama3.1 generates PM₂.₅ emissions equivalent to over 10,000 round-trip automobile journeys between Los Angeles and New York. Health burdens exhibit pronounced spatial heterogeneity—up to a 200-fold disparity across U.S. census tracts—disproportionately affecting marginalized communities. By 2030, AI-related datacenter operations in the U.S. are projected to incur annual health costs exceeding $20 billion. We propose mandatory disclosure standards for AI health externalities and a health-centered governance framework to advance equitable, sustainable AI development.

13 citations1 influentialRead paper

Extending the Control Plane of Container Orchestrators for I/O Virtualization

Nov 01, 2020International Workshop on Containers and New Orchestration Paradigms for Isolated Environments in HPC

Existing container orchestrators lack dynamic SR-IOV virtual network device configuration aligned with application bandwidth requirements and offer insufficient fine-grained control over RDMA virtualization interfaces. To address this, we propose ConRDMA—a novel SR-IOV extension control plane that integrates container orchestration APIs, a customized RDMA driver, and a bandwidth-aware scheduling algorithm. ConRDMA enables configurable, fine-grained allocation of container-level RDMA virtual devices and optimizes cross-node communication paths. It realizes bandwidth-driven dynamic resource scheduling while preserving RDMA’s high-performance communication capabilities and significantly reducing virtualization overhead. Experimental evaluation demonstrates that ConRDMA improves RDMA bandwidth utilization by over 35% and supports optimal node scheduling tailored to application-specific communication demands.

3 citations1 influentialRead paper

On-the-Fly VLA Adaptation via Test-Time Reinforcement Learning

Jan 11, 2026arXiv.org

This work proposes TT-VLA, a novel framework that introduces test-time reinforcement learning to vision-language-action (VLA) models, enabling online adaptation during deployment without requiring retraining. Existing VLA models lack the ability to adapt at test time, limiting their robustness in dynamic environments. TT-VLA addresses this by fine-tuning policies at inference using task-progress signals, integrating a dense reward mechanism with prior-preserving techniques to maintain stable and effective behavior. Experiments demonstrate that the method significantly improves task success rates, policy stability, and adaptability to unseen environmental dynamics in both simulated and real-world settings, thereby enhancing the practical deployability of VLA models.

1 citationsRead paper

Do We Really Need to Design New Byzantine-robust Aggregation Rules?

Jan 29, 2025

In federated learning, Byzantine clients launching poisoning attacks can severely degrade the robustness of existing aggregation rules. To address this, we propose FoundationFL—a framework that preserves standard robust aggregators (e.g., Trimmed-mean, Median) without modifying their logic; instead, the server generates synthetic model updates, which are jointly aggregated with clients’ local updates. We provide the first theoretical proof that enhancing input quality alone—without designing new aggregation rules—significantly improves Byzantine resilience of classical robust aggregators. FoundationFL guarantees convergence under Byzantine threats and empirically demonstrates substantial improvements in poisoning resistance across multiple real-world datasets, while maintaining high model accuracy and low communication overhead. The framework thus achieves strong effectiveness, generalizability, and practicality.

1 citationsRead paper
Recent publications

Latest Papers