A retrieval-augmented generation system can be poisoned with reward-optimized biased documents and vector-space manipulation to substantially increase biased LLM outputs.
Mitigating Bias in RAG: Controlling the Embedder
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
In retrieval augmented generation (RAG) systems, each individual component -- the LLM, embedder, and corpus -- could introduce biases in the form of skews towards outputting certain perspectives or identities. In this work, we study the conflict between biases of each component and their relationship to the overall bias of the RAG system, which we call bias conflict. Examining both gender and political biases as case studies, we show that bias conflict can be characterized through a linear relationship among components despite its complexity in 6 different LLMs. Through comprehensive fine-tuning experiments creating 120 differently biased embedders, we demonstrate how to control bias while maintaining utility and reveal the importance of reverse-biasing the embedder to mitigate bias in the overall system. Additionally, we find that LLMs and tasks exhibit varying sensitivities to the embedder bias, a crucial factor to consider for debiasing. Our results underscore that a fair RAG system can be better achieved by carefully controlling the bias of the embedder rather than increasing its fairness.
citation-role summary
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
REJECT 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Bias Amplification in RAG: Poisoning Knowledge Retrieval to Steer LLMs
A retrieval-augmented generation system can be poisoned with reward-optimized biased documents and vector-space manipulation to substantially increase biased LLM outputs.