Rudinger Developing Large Language Models with Commonsense Reasoning

With the use of large language models (LLMs) like OpenAI's ChatGPT growing exponentially, the technology’s quirks and limitations have become strikingly more evident.

LLMs can falter, for example, when prompted to make simple predictions about everyday situations like cooking a meal or riding in a car, thus revealing significant gaps in judgement and commonsense reasoning.

Sometimes these garbled or incorrect outputs from LLMs reflect social stereotypes and cultural assumptions that limit the usefulness of the technology for certain populations at best, and at worst, cause active harm.

With funding from the National Science Foundation (NSF), Assistant Professor of Computer Science Rachel Rudinger aims to address these issues—focusing her research on enhancing the robustness, fairness and cultural adaptability of commonsense reasoning in LLMs.

Her work is supported by an NSF Faculty Early Career Development Program (CAREER) award, given to junior faculty that exemplify the role of teacher-scholars through their outstanding research and scholarship.

The NSF funding, expected to total just under $600,000 over the next five years, will significantly advance Rudinger’s work focused on the interplay between language and computational commonsense reasoning.

“These problems are real, but are often difficult to measure, so a major objective for us is to advance the science of LLM evaluation in these areas,” says Rudinger, who has a joint appointment in the University of Maryland Institute for Advanced Computer Studies. “The long-term vision is to work toward language technologies equipped with nuanced and unprejudiced reasoning skills that can adapt to meet the needs of a wider set of users across cultures.”

The project aims to develop scientific methods to measure and improve the ability of LLMs to reason correctly about common situations in a way that’s fair and unbiased, and then adapt these reasoning abilities across specific cultural contexts.

By measuring these fundamental capabilities of LLMs, Rudinger says, researchers can better understand and mitigate the risks of applying this technology when used in high-stakes settings like education or health care.

The NSF funding includes provisions for developing a new curriculum and teaching materials that integrate Rudinger’s research findings into classroom instruction modules. This is intended to nurture the next generation of scientists and engineers, equipping them with the knowledge to tackle complex problems in artificial intelligence.

“I am simultaneously thrilled and humbled that NSF has chosen to support my research vision,” says Rudinger, who is also a member of the university’s Computational Linguistics and Information Processing (CLIP) Lab. “I am fortunate to work with a group of outstanding Ph.D. students here at the University of Maryland, and it is gratifying to be able to support them through an award that could not have materialized without their hard work.”

***

“Robust, Fair, and Culturally Aware Commonsense Reasoning in Natural Language” is supported by NSF grant #2339746 from NSF’s Division of Information and Intelligent Systems.

PI: Rachel Rudinger, an assistant professor of computer science with an appointment in the University of Maryland Institute for Advanced Computer Studies.

About the CAREER award: The Faculty Early Career Development (CAREER) Program is an NSF activity that offers the foundation’s most prestigious awards in support of junior faculty who exemplify the role of teacher-scholars through outstanding research, excellent education and the integration of education and research within the context of the mission of their organization.