A study examines how Retrieval-Augmented Generation (RAG) frameworks transmit ideological bias from retrieved materials into large language model (LLM) responses. The research identifies that sampling temperature significantly influences the strength of this ideological transfer.
- Researchers applied Lexical Multidimensional Analysis to a corpus of 1,117 COVID-19 treatment articles to identify three ideological discourses used as external knowledge.
- LLMs were tested on ideological questions at varying sampling temperatures, with outputs assessed for semantic and lexical similarity to reference texts.
- Discoursive alignment between generated answers and reference texts peaks at moderate temperatures, balancing stochasticity with retrieval grounding.
- Alignment drops at low temperatures, indicating that overly deterministic sampling suppresses the transfer of ideological discourse.
The findings highlight that RAG systems are prone to amplifying or suppressing ideological positions based on their retrieval sources, with temperature serving as a critical control factor for discourse alignment.