Homogeneity Bias in Open-Weight LLMs Is Robust to Decoding Hyperparameters

A systematic study of seven open-weight large language models finds that the tendency to portray marginalized groups as more internally similar than dominant groups — known as homogeneity bias — persists across a wide range of decoding settings, challenging the assumption that such bias is merely an artifact of inference hyperparameters [1][2]. The paper, posted on the arXiv preprint repository and last revised on 24 June 2026, was authored by Messi H.J. Lee [1][2]. It examines seven instruction-tuned models ranging from 7 billion to 20 billion parameters [1][2]. The researchers mapped homogeneity bias across a 5-by-5 grid of temperature and top-p values, creating 20 distinct hyperparameter configurations per model [1][2]. Hispanic and Asian Americans were portrayed as more homogeneous than White Americans in at least 18 of the 20 configurations across six of the seven models tested, including at extreme sampling settings [1][2]. A conservative cell-level re-analysis confirmed that the Hispanic and Asian homogeneity signals were robust, while weaker signals for African American and gender bias largely did not survive the same scrutiny [1][2]. The study also tested a names-based paradigm, in which group identity was signaled through racially distinctive surnames rather than explicit labels. That approach corroborated the Hispanic and Asian homogeneity bias, but produced a reversal for Black-coded surnames, which elicited less homogeneous outputs than White-coded names in every model examined [1][2]. The finding underscores that how group identity is operationalized can shape both the direction and the magnitude of the bias that surfaces [1][2]. arXiv, where the paper appears, is an open-access repository of electronic preprints that are moderated but not peer-reviewed [6]. It was launched in 1991 and now receives roughly 24,000 submissions per month [6]. The platform also hosts arXivLabs, a framework that allows community collaborators to build experimental tools on top of the repository, though the Labs program is currently pausing new proposals while the development team focuses on migrating systems to the cloud [3][4].