SBIR Phase I: RELAI: Enhancing Reliability of Large Language Models

Relai, Inc., Bethesda MD

Investigators

Abstract

The broader/commercial impact of this Small Business Innovation Research (SBIR) Phase I project are significant and multifaceted. By improving the reliability of advanced artificial intelligence (AI) models such as Large Language Models (LLMs), this project will contribute to safer and more effective AI implementations across various industries, thereby enhancing economic competitiveness and fostering public trust in AI technologies. This project not only contributes to the foundational understanding of LLM reliability but also provides practical solutions that can be widely adopted in the industry. Additionally, the enhancements in AI reliability will have far-reaching impacts on sectors like healthcare, finance, and customer service, where accurate and unbiased decision-making is crucial. For instance, in healthcare, reliable LLMs can lead to better diagnostic tools and personalized treatment plans, ultimately improving patient outcomes and reducing healthcare costs. In finance, these models can enhance risk assessment and fraud detection, providing more secure and efficient financial services. The project’s integration with academic and industry partners ensures that the developed technologies are not only cutting-edge but also grounded in real-world applicability. Furthermore, the project includes a strong educational component aimed at training the next generation of AI practitioners in ethical AI practices. This Small Business Innovation Research (SBIR) Phase I project introduces innovative methodologies to enhance the reliability of large language models (LLMs). In particular, the project presents methodologies to inspect and mitigate jailbreaking issues of LLMs, where adversarial prompts can circumvent their alignment, methodologies to inspect and mitigate LLM hallucinations, where models can generate non-factual responses, and methodologies to inspect and mitigate LLM biases. In particular, the development of methodologies to inspect and mitigate jailbreaking in LLMs represents a significant advancement in adversarial machine learning. This work not only identifies the vulnerabilities in current LLMs but also proposes robust countermeasures to fortify these models against sophisticated attacks. Moreover, the methodologies to inspect and mitigate hallucinations in LLMs involves sophisticated analysis of model outputs to identify when and why hallucinations occur, providing deeper insights into the internal workings of LLMs. Finally, addressing biases in LLMs is a critical component of ensuring ethical and fair artificial intelligence (AI). By integrating these advanced tools into a comprehensive, user-friendly, and unified platform, this project establishes a new benchmark for the development and deployment of reliable AI applications. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.