How Temperature Shapes Ideological Discourse in Retrieval-Augmented Generation?

2026-07-13

AI Summary

This study investigates how retrieval-augmented generation (RAG) systems propagate, amplify, or suppress ideologically charged discourse when incorporating external knowledge with explicit ideological stances, and for the first time systematically reveals the modulating role of sampling temperature in this process. Leveraging a corpus of 1,117 articles on COVID-19 treatments, the research identifies three categories of ideological discourse as retrieval sources and evaluates generative alignment across multiple large language models under varying temperatures using lexical multidimensional analysis (LMDA) and semantic similarity metrics. The findings demonstrate that RAG systems readily transmit ideological discourse from retrieved content into model outputs, with alignment strength significantly influenced by temperature: maximal alignment occurs at moderate temperatures, whereas low temperatures suppress discourse propagation due to excessive output determinism.

Abstract

Retrieval-Augmented Generation (RAG) has been increasingly adopted to reduce hallucinations and strengthen the factual grounding of large language models (LLMs). While robustness to errors in the retrieval process has been explored, the impact of ideological bias on LLM outputs has been overlooked. For instance, if the retrieved material contains ideological positions, the RAG may transmit, amplify, or suppress such ideological discourses in its outputs. In this study, we address this issue by examining the influence of the RAG framework, comprising ideological discourses, in LLM-generated answers. To this end, we applied Lexical Multidimensional Analysis (LMDA) on a corpus of 1,117 COVID-19 treatment articles, identifying three ideological discourses. This corpus is then used as the external knowledge source for the RAG. We assessed several LLMs by having the models answer ideological questions at different sampling temperatures. The generated texts were assessed semantically and lexically based on their similarities with ideological reference texts. Our findings show that the RAG framework is prone to transferring ideological discourses into LLM responses, with sampling temperature having a measurable impact on the strength of this transfer. Discoursive alignment between generated answers and the reference text is highest at moderate temperatures, where models balance stochasticity with retrieval grounding, and drops at low temperatures, indicating that overly deterministic sampling suppresses discourse transfer.

Problem

• Retrieval-Augmented Generation
• ideological bias
• sampling temperature
• large language models
• ideological discourse

Innovation

• Retrieval-Augmented Generation
• ideological bias
• sampling temperature
• Lexical Multidimensional Analysis
• discourse transfer