Principles of Augmented Language Models: Accelerating Skill Emergence, Increasing Capacity, Improving Reliability towards General Purpose AI
Abstract
Approved for Public ReleaseModern machine learning, with Transformers (TFs) at its forefront, has exhibited impressiveperformance in Natural Language Processing (NLP) and Computer Vision (CV). Large LanguageModels (LLMs) like GPT-3/4, PaLM, and LaMDA have unlocked emergent general-purposeskills, demonstrating significant capability for a diverse range of tasks. However, despite theseadvancements, limitations persist related to computational capacity, reliability, and the emergenceof skills. This highlights the necessity for a thorough exploration of the underlying principles ofthese models and the development of strategies to optimize their potential.The grand endeavor of this ONR project is to unravel the complexities of augmented languagemodels, optimize their emergent general-purpose skills, enhance their capacity for reasoning, andimprove their reliability. We aim to drive this field towards the realization of General Purpose AI,by conducting a comprehensive analysis of Large Language Models and their emergent abilities.Our research plan focuses on four central objectives. First, we investigate the foundations ofthe emergence of general-purpose skills in LLMs, exploring the driving factors behind theirdevelopment. Then, we delve into understanding the principles and challenges associated within-context learning (ICL) and compositional reasoning. We also examine the potential ofaugmenting LLMs with loops, external memory, and add-on skills to enhance their capacityfor problem-solving and reasoning. Concurrently, we tackle the issue of detecting and preventinghallucinations in LLMs to increase their reliability in critical applications.Our proposed research is anticipated to be transformational in the field of AI, significantlyimpacting the design of future machine learning systems. If successful, this project is poised tocreate a new understanding of emergent abilities in LLMs, enabling the development of moreefficient and reliable AI systems. We are motivated by a vision of pushing the boundaries of AI,contributing substantially to the pursuit of developing general-purpose AI systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 13, 2023
- Source ID
- N000142312848
Entities
People
- Dimitrios Papailiopoulos
Organizations
- Office of Naval Research
- United States Navy
- University of Wisconsin System