Stony Brook Awards Grants to Bridge AI Research and Real-World Application

Stony Brook University’s AI Innovation Institute has awarded small grants to ten interdisciplinary research projects, selected from a pool of sixty-three applications. Lav Varshney, the institute’s inaugural director, emphasized that these initiatives represent a crucial intersection of technological innovation and societal application. The funding aims to move beyond theoretical algorithm development to address tangible challenges in fields ranging from epilepsy treatment to the study of mammalian evolution. This approach reflects a broader institutional strategy to ensure that advances in foundational AI technology are translated into practical tools that benefit society.

The grant program is structured around three distinct tracks designed to foster comprehensive progress in the field. The first track, Innovation in AI, supports research into new algorithms, mathematical foundations, and ethical frameworks. The second, Diffusion of AI, focuses on integrating these technologies into industrial, societal, or scholarly sectors where they have not previously been used. The third track targets the creation of high-quality datasets and benchmark tasks that respect the core disciplines they serve, such as health sciences, journalism, and physical sciences. This tripartite structure ensures that the funded work contributes to the entire ecosystem of AI development, not just isolated technical improvements.

Ensuring Safety in Continual Learning

One of the selected projects, led by Associate Professor Jian Li, addresses a critical reliability issue in large language models. As these models are frequently updated to meet new safety policies or user needs, there is a risk that improving one function can inadvertently break another. Current alignment methods lack guarantees against these regressions, which can erode user trust if an assistant fails a safety check it previously passed. Li’s project proposes a framework called Regression-Gated Continual Alignment. This system treats model updates as a continual learning process constrained by explicit regression gates. Before any update is released, it must pass a compact suite of tests for safety, policy, and tool validity. The project aims to provide formal metrics for regression risk and mathematical guarantees that will make AI agents more stable and trustworthy for end users.

Detecting Hidden Consciousness in Patients

Another funded initiative, led by Assistant Professor Akshat Dave and Co-Principal Investigator Ulas Sunar, tackles a severe diagnostic gap in neurocritical care. Approximately forty percent of patients diagnosed with a vegetative state after traumatic brain injury are actually conscious, a condition that standard behavioral assessments often miss. These standard evaluations are brief snapshots that cannot detect hidden awareness, potentially leading to the premature withdrawal of life support. The project seeks to develop an automated, non-invasive monitoring system that combines electroencephalography with diffuse correlation spectroscopy. By using near-infrared light to measure deep cerebral blood flow, the system can continuously record brain activity across various states of consciousness. An AI model will fuse these data streams to provide objective, real-time indicators of awareness, offering a more reliable tool for medical teams to make critical care decisions.

Balancing Innovation with Practical Limits

While these grants promise significant advancements, they also highlight the inherent trade-offs in deploying AI in sensitive environments. The project by Li acknowledges that adding safety constraints can limit the flexibility of model updates, a necessary compromise to prevent regressions. Similarly, Dave’s project relies on complex sensor fusion, which requires rigorous calibration and interpretation to avoid false positives in a high-stakes medical setting. As reported by GN technics/ai (en-US), these efforts underscore that the value of AI lies not just in its computational power, but in its ability to operate reliably within the complex, often messy realities of human health and society. The success of these projects will depend on their ability to balance technical rigor with ethical responsibility.