Open-weight LLMs show robust homogeneity bias across decoding settings

A new study published on arXiv investigates homogeneity bias in open-weight large language models (LLMs). Researchers found that models tend to portray marginalized groups as more internally similar than dominant groups, and this bias remains consistent even when adjusting decoding hyperparameters. The study tested seven open-weight LLMs and observed that Hispanic and Asian Americans were consistently depicted as more homogeneous across various sampling settings. The research also highlighted that the method used to signal group identity significantly impacts the observed biases, with a name-based approach revealing a reversal in bias direction for Black-coded surnames compared to explicit labels. AI

IMPACT Highlights potential biases in open-weight LLMs that could affect fairness and representation in AI applications.

RANK_REASON Academic paper detailing research findings on LLM bias. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Asian Americans
• Large language models
• Open-weight LLMs
• White Americans

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →