Messi H. J. Lee

Messi H. J. Lee, Calvin K. Lai

Homogeneity bias in Large Language Models (LLMs) refers to their tendency to homogenize the representations of some groups compared to others. Previous studies documenting this bias have predominantly used encoder models, which may have inadvertently introduced biases. To address this limitation, we prompted GPT-4 to generate single word/expression completions associated with 18 situation cues-specific, measurable elements of environments that influence how individuals perceive situations and compared the variability of these completions using probability of differentiation. This approach directly assessed homogeneity bias from the model's outputs, bypassing encoder models. Across five studies, we find that homogeneity bias is highly volatile across situation cues and writing prompts, suggesting that the bias observed in past work may reflect those within encoder models rather than LLMs. Furthermore, we find that homogeneity bias in LLMs is brittle, as even minor and arbitrary changes in prompts can significantly alter the expression of biases. Future work should further explore how variations in syntactic features and topic choices in longer text generations influence homogeneity bias in LLMs.

Messi H. J. Lee

As autonomous AI agents are deployed in persistent, interacting networks -- coordinating tasks, routing resources, and accumulating reputational histories -- the social dynamics that emerge will determine who receives opportunity and who does not, at scales no human institution can supervise. We ran a controlled multi-agent simulation in which instruction-tuned language model agents interacted across 500 turns under three conditions manipulating group label salience and resource scarcity, across six model families with 20 seeds each. When group labels were visible, we observed in-group trust bias, action homophily, and network assortativity -- all absent when labels were hidden -- a pattern structurally consistent with salience-dependence in human social psychology. This discrimination was invisible to standard action-log audits: bias operated entirely through who received each action, not what actions were chosen, with action-type distributions showing no increase in negative actions across conditions. Per-turn in-group versus out-group differentials of 5 to 16 percentage points were statistically significant for all six models (Wilcoxon signed-rank, all Benjamini-Hochberg-corrected p < 0.001), establishing group-contingent targeting as a robust property of instruction-tuned language models across architectures and training regimes. Compounded through 500 turns of reciprocation, these differentials accumulated into in-group trust biases of +0.014 to +0.100 (d = 0.84-4.52) -- illustrating how modest per-interaction targeting propagates into structural inequality in persistent networks.

Messi H. J. Lee

Vision-Language Models trained on massive collections of human-generated data often reproduce and amplify societal stereotypes. One critical form of stereotyping reproduced by these models is homogeneity bias-the tendency to represent certain groups as more homogeneous than others. We investigate how this bias responds to hyperparameter adjustments in GPT-4, specifically examining sampling temperature and top p which control the randomness of model outputs. By generating stories about individuals from different racial and gender groups and comparing their similarities using vector representations, we assess both bias robustness and its relationship with hyperparameter values. We find that (1) homogeneity bias persists across most hyperparameter configurations, with Black Americans and women being represented more homogeneously than White Americans and men, (2) the relationship between hyperparameters and group representations shows unexpected non-linear patterns, particularly at extreme values, and (3) hyperparameter adjustments affect racial and gender homogeneity bias differently-while increasing temperature or decreasing top p can reduce racial homogeneity bias, these changes show different effects on gender homogeneity bias. Our findings suggest that while hyperparameter tuning may mitigate certain biases to some extent, it cannot serve as a universal solution for addressing homogeneity bias across different social group dimensions.

Messi H. J. Lee, Calvin K. Lai

Implicit biases refer to automatic mental processes that shape perceptions, judgments, and behaviors. Previous research on "implicit bias'' in LLMs focused primarily on outputs rather than the processes underlying the outputs. We present the Reasoning Model Implicit Association Test (RM-IAT) to study implicit bias-like processing in reasoning models, which are LLMs that use step-by-step reasoning for complex tasks. Using RM-IAT, we find that reasoning models like o3-mini, DeepSeek-R1, gpt-oss-20b, and Qwen-3 8B consistently expend more reasoning tokens on association-incompatible tasks than association-compatible tasks, suggesting greater computational effort when processing counter-stereotypical information. In contrast, Claude 3.7 Sonnet exhibited reversed or inconsistent patterns, likely due to embedded safety mechanisms that flagged or rejected socially sensitive associations. These divergent behaviors highlight important differences in how alignment and safety processes shape model reasoning. As reasoning models become increasingly integrated into real-world decision-making, understanding their implicit bias-like patterns and how alignment methods influence them is crucial for ensuring fair and trustworthy AI systems.

Messi H. J. Lee, Soyeon Jeon

Homogeneity bias is one form of stereotyping in AI models where certain groups are represented as more similar to each other than other groups. This bias is a major obstacle to creating equitable language technologies. We test whether the bias is driven by systematic differences in token-sampling uncertainty across six large language models. While we observe the presence of homogeneity bias using sentence similarity, we find very little difference in token sampling uncertainty across groups. This finding elucidates why temperature-based sampling adjustments fail to mitigate homogeneity bias. It suggests researchers should prioritize interventions targeting representation learning mechanisms and training corpus composition rather than inference-time output manipulations.