How Temperature Shapes Ideological Discourse in Retrieval-Augmented Generation?

This study investigates how Retrieval-Augmented Generation (RAG) systems propagate ideological biases from external knowledge sources into LLM-generated responses. Specifically, the authors examine how sampling temperature—a parameter controlling the randomness of text generation—modulates the strength of this ideological transfer, testing whether models amplify, transmit, or suppress the ideological framing present in retrieved documents.

The researchers constructed a corpus of 1,117 scientific articles regarding COVID-19 treatments, categorized into "endorsed" (standard scientific) and "controversial" (non-standard) discourses. Using Lexical Multidimensional Analysis (LMDA), they identified three distinct ideological dimensions within this corpus. They then evaluated several LLMs using these documents as the external knowledge base for RAG. The models were prompted to answer ideological questions across a range of sampling temperatures (0.1 to 0.9) to measure how the discursive alignment between the generated answers and the reference texts shifted based on decoding randomness.

The study demonstrates that RAG systems are highly susceptible to transferring ideological patterns from retrieved documents into their outputs. The researchers found that sampling temperature acts as a critical control for this phenomenon: discursive alignment is highest at moderate temperatures, where the model balances stochasticity with the grounding provided by the retrieved context. Conversely, at low temperatures, the model's output becomes overly deterministic, which effectively suppresses the transfer of the nuanced ideological discourse found in the source material.