Reproduce “Why Uncertainty Estimation Methods Fall Short in RAG: An Axiomatic Analysis” on OpenResearch

About this paper

Heydar Soudani, Evangelos Kanoulas, Faegheh Hasibi

Existing uncertainty estimation methods for Large Language Models demonstrate inconsistent performance and a bias towards lower uncertainty in Retrieval-Augmented Generation (RAG) contexts, often failing to accurately reflect response correctness. An axiomatic framework identifies these deficiencies, and a calibration function improves method reliability by enforcing desired properties, leading to enhanced correlation between uncertainty estimates and actual response accuracy.

Key insights

• Existing UE methods exhibit a systematic bias, consistently reducing uncertainty scores in RAG setups regardless of context relevance, which can lead to a false sense of confidence.
• Current UE methods largely fail to satisfy critical axioms, particularly those requiring uncertainty to increase when context contradicts a response or when the LLM switches to an incorrect answer after RAG.
• The proposed axiomatic framework effectively diagnoses shortcomings in UE for RAG and offers a principled guide for evaluating and developing more reliable uncertainty mechanisms.

Summary by alphaXiv.