Authors:

Major: Computer Science

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Understanding Morality-Related Biases in LLMs: A Cross-Demographic Fine- Grained Analysis The prevalence of large language models (LLMs) in high-impact domains raises important concerns regarding moral reasoning and value alignment. As LLMs are trained on large- scale corpora, human beliefs, societal norms, and prejudices inherently get captured in model representations. In this paper, a fine-grained analysis of morality-related biases in LLMs is explored across multiple demographic characteristics (e.g., age, race, ability). Existing work on social and moral bias in LLMs emphasizes gender bias, highlighting systematic differences in moral judgments when presented with male versus female subjects. While these findings highlight the importance of evaluating moral bias, moral reasoning is shaped by a variety of social attributes. Therefore, it is important to understand broader and more nuanced patterns of social bias. To investigate this issue, we construct parallel moral scenarios, such that demographic characteristics are systematically perturbed without altering context and narrative structure. This allows for controlled comparisons of model-generated moral judgments and accompanying rationales across demographic conditions. We then examine how distributions of moral opinions shift across population groups. This work aims to provide a more comprehensive understanding of how morality-related biases are present in LLMs across diverse demographic contexts. By broadening the scope of moral bias evaluation beyond gender, this study contributes to ongoing efforts to assess and improve the fairness, robustness, and ethical alignment of LLMs deployed in morally sensitive applications. Poster Session 3 1:00 PM-2:00 PM CT Room C Poster #17: Roman Parker Major: Computer Science Faculty Advisors: Dr. Tracy Hammond, Dr. Adam Kolasinski Adaptive Factor Allocation for Macroeconomic State-Aware Factor Investing This project develops a Dynamic Factor Allocation Model that combines machine learning factor discovery with macroeconomic regime detection to improve equity portfolio performance. Traditional factor models assume that style factors like value, momentum, and quality have consistent behavior over large time scales, but research shows their performance often varies with changing economic conditions. Using publicly accessible equity data, we apply Instrumented PCA (IPCA) and similar techniques to locate latent & interpretable factors to capture time-varying loadings. At the same time, we use Bayesian changepoint detection and hidden Markov models to spot larger shifts in macroeconomic regimes in near real time. These regime factor premia are applied via a Black-Litterman allocation framework, adjusting factor exposures as regime shifts occur. This provides an algorithm for factor investing capable of reacting to macroeconomic regimes without sacrificing the interpretability and stability that makes traditional style factors effective. Poster Session 3 1:00 PM-2:00 PM CT Room C Poster #18: Audrey Hillam Majors: International Affairs, Modern Languages

Source:

Texas A&M University / 2026

Topics:

No topics listed

Co-authors:

Major: Computer Science