Search papers, labs, and topics across Lattice.
José Victor de Souza, Institut national de la recherche scientifique, Québec, Montréal, Canada
2
0
3
Ideological biases in RAG can significantly shape LLM outputs, with moderate sampling temperatures amplifying discourse alignment.
RAG systems readily absorb and amplify ideological biases present in retrieved documents, even more so when prompts explicitly describe the ideological dimensions at play.
