arXiv:2501.02211v2 Announce Type: replace-cross Abstract: Large language models (LLMs) reproduce homogeneity bias -- the tendency to portray marginalized groups as more internally similar than dominant groups -- but whether this bias is stable or an artifact of inference settings has only been studied in single proprietary models. We map homogeneity bias across a 5x5 temperature-by-top-p grid in seven open-weight instruction-tuned LLMs (7-20B parameters). Hispanic and Asian Americans are portrayed as more homogeneous than White Americans in at least 18 of 20 hyperparameter configurations acros
This study is published as open-weight LLMs become more widely used, necessitating a deeper understanding of their inherent biases across various configurations.
The persistence of homogeneity bias in open-weight LLMs, regardless of typical decoding customizations, indicates a foundational issue in model training or architecture that impacts fairness and representation.
This research suggests that bias mitigation efforts will need to go beyond simple inference parameter tuning and target more fundamental aspects of model development and data curation.
- · AI ethics researchers
- · Organizations prioritizing fair AI
- · Specialized bias mitigation platforms
- · Developers relying solely on decoding parameters for bias control
- · Users impacted by stereotypical representations
- · Generic LLM deployment strategies
Increased scrutiny and research into LLM training data and pre-alignment strategies.
Development of new architectural or fine-tuning approaches specifically designed to counteract homogeneity bias at its source.
Potential for regulatory frameworks to mandate bias testing and transparency in large language models preceding deployment.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG