PulseAugur
EN
LIVE 08:41:24

RAG systems transmit ideological bias, influenced by sampling temperature

By PulseAugur Editorial · [2 sources] ·

A new study published on arXiv investigates how the sampling temperature parameter in Retrieval-Augmented Generation (RAG) systems affects the transmission and amplification of ideological biases present in retrieved information. Researchers analyzed 1,117 COVID-19 treatment articles to identify three distinct ideological discourses, which were then used as an external knowledge source for RAG models. The findings indicate that RAG frameworks can indeed transfer ideological discourses into LLM responses, with moderate temperatures showing the highest alignment between generated answers and reference texts, while lower temperatures suppress this transfer. AI

IMPACT This research highlights a potential vulnerability in RAG systems, suggesting that careful tuning of sampling temperature may be necessary to mitigate the spread of ideological bias in AI-generated content.

RANK_REASON The cluster contains an academic paper detailing research findings on LLM behavior.

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

RAG systems transmit ideological bias, influenced by sampling temperature

COVERAGE [2]

  1. arXiv cs.CL TIER_1 English(EN) · Elmira Salari, Hazem Amamou, Jos\'e Victor de Souza, Shruti Kshirsagar, Maria Nunes Delfino, Anderson Avila ·

    How Temperature Shapes Ideological Discourse in Retrieval-Augmented Generation?

    arXiv:2607.11783v1 Announce Type: new 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,…

  2. arXiv cs.CL TIER_1 English(EN) · Anderson Avila ·

    How Temperature Shapes Ideological Discourse in Retrieval-Augmented Generation?

    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 h…

RELATED ENTITIES

RELATED TOPICS