��������

[Defense] A Human-Centered Evaluation of Rationality and Legal Reasoning in Large Language Models (LLMs): An Audit Guided by the Cattell-Horn-Carroll Theory

Friday, December 6, 2024

11:00 am - 12:30 pm

In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
Dana Alsagheer

will defend her proposal
A Human-Centered Evaluation of Rationality and Legal Reasoning in Large Language Models (LLMs): An Audit Guided by the Cattell-Horn-Carroll Theory

Abstract

Large language models (LLMs), such as OpenAI’s GPT-4, are increasingly deployed in high-stakes domains like law, healthcare, scientific research, and text generation. Despite their impressive capabilities, these models function as “black-box” systems, prompting significant concerns around transparency, reliability, and ethical integrity. Key issues include biases embedded in training data, inconsistencies in reasoning, limited reproducibility, and an inflexible approach to adapting new information. These limitations present notable risks in domains where fairness, accountability, and transparency are paramount. Traditional evaluations tend to focus on surface-level accuracy, often overlooking the complex cognitive mechanisms underlying LLM responses and the dynamic shifts in user beliefs that develop through continued interaction. This study explores the real implications of LLMs passing domain-specific exams, particularly regarding cognitive robustness and applicability to real-world scenarios. To achieve a thorough evaluation, we introduce a two-way auditing framework aimed at comprehensive, accurate, and reliable assessments of LLM capabilities. The first component of this framework is the Human Generalization Function (HGF). In this interactive assessment, users use the model to evaluate its reasoning and coherence across multiple interactions. The HGF tracks shifts in user perception, observing how users adjust their understanding of the model’s abilities over time based on its responses, thus providing insights into the model’s perceived trustworthiness and consistency. The second component involves simulated exams administered by domain experts to assess competencies in narrow domains, such as legal reasoning. These exams offer a transparent, standardized metric for evaluating LLM performance, providing a clear benchmark to compare model capabilities against established human standards in specialized fields. By integrating these methods, our framework surpasses basic accuracy metrics to investigate LLM performance across both broad and narrow cognitive abilities, assessing whether improvements in specific skills indicate isolated enhancement or suggest broader generalization in line with the Cattell-Horn-Carroll (CHC) theory of intelligence. To support this investigation, we audit five distinct LLMs and offer a comparative analysis highlighting each model’s strengths and limitations in various high-stakes applications. This detailed evaluation encompasses domains requiring precision, ethical reasoning, and logical consistency, advancing our understanding of how these models align—or fail to align—with human cognitive standards in critical decision-making contexts. Ultimately, our framework seeks to establish a new standard for ethical AI governance by merging rigorous cognitive assessments with dynamic user feedback. This approach enables a transparent, reliable, and human-aligned evaluation of LLMs, illuminating their capabilities and limitations. In doing so, we provide essential guidance for the future development of AI that prioritizes transparency, trustworthiness, and alignment with human values—especially in high-stakes areas where accuracy, fairness, and accountability are crucial.

Friday, December 6, 2024
11:00 AM - 12:30 PM

PGH 591

Dr. Weidong (Larry) Shi, proposal advisor

Faculty, students, and the general public are invited.