New method tackles topic-level fairness in retrieval-augmented generation

Researchers have introduced a new approach to evaluate and improve topic-level fairness in retrieval-augmented generation (RAG) systems. Current methods often aggregate fairness metrics across all queries, which can mask significant disparities within specific topics. The proposed method adapts rerankers for each topic individually, using preference pairs derived from the RAG system's own generated output to guide the adaptation. This technique has demonstrated a substantial reduction in topic-level disparities on benchmark datasets like RoMQA and TREC Fair Ranking, while maintaining accuracy. AI

IMPACT Addresses a critical gap in AI fairness evaluation, potentially leading to more equitable AI systems across diverse applications.

RANK_REASON Academic paper detailing a new method for evaluating and improving AI model fairness. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Bradley--Terry model
• retrieval-augmented generation
• RoMQA
• TREC Fair Ranking

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →