Rose Yu, an associate professor from University of California San Diego’s Department of Computer Science and Engineering (CSE), has been awarded the DARPA Young Faculty Award for research aimed at improving the ability of large language models (LLMs) to solve complex real-world tasks with unseen scenarios.
The Defense Advanced Research Projects Agency, or DARPA, partners with academic and industry leaders to create breakthrough technologies and capabilities for national security.
Yu’s proposal, “Grounding LLMs with Physical Laws,” aligns scientific research with DARPA’s explicit mission. Her proposal introduces a novel, hybrid artificial intelligence (AI) approach for training deep learning models – one that improves the trustworthiness of LLMs to both boost scientific productivity and support military objectives.
Existing AI models can only learn statistical patterns from massive data, and they struggle to generalize to unseen conditions. In contrast, Yu’s framework integrates physical laws into deep neural networks in a principled manner, effectively grounding LLMs.
The project will lead to innovations in equivariant deep learning, constrained energy-based models, and fine-tuning LLMs with reward models. Breakthroughs such as these will enable LLMs to generate creative content with guaranteed physical consistency and scientific factuality, while expanding their capabilities to model diverse environments and heterogeneous conditions.
The research will also have a direct impact on national security applications, including safe-operation of autonomous weapons in the battlefield, optimal allocation of military resources, and automated document processing and intelligence gathering for real-time mission planning and control.
By Kimberley Clementi