Diagnosing LLM Arbitration Behavior over Pre-evidence Epistemic States in RAG-based Fact-Checking
Summary
A new diagnostic testbed, \textsc{PAVE} (Prior-Aware Verifier Evaluation), has been introduced to analyze Large Language Model (LLM) arbitration behavior in RAG-based fact-checking. LLMs used as verifiers often exhibit pre-evidence tendencies from their parametric knowledge, which can conflict with retrieved contextual evidence. Existing evaluation frameworks do not adequately characterize this prior-context discrepancy or measure how LLMs arbitrate between these signals. \textsc{PAVE} stratifies an LLM verifier into four epistemic states based on its pre-evidence prior's correctness and confidence, then evaluates its ability to persist in correct priors under misleading evidence or correct wrong priors with accurate evidence. Experiments across seven LLMs revealed unreliable and highly model-dependent prior-context arbitration. Based on these findings, a lightweight JSD-based test-time arbitration method is proposed, which enhances factual reliability without modifying the underlying model, achieving competitive performance across diverse LLM families.
Key takeaway
For Machine Learning Engineers deploying LLMs in RAG-based fact-checking, you must critically evaluate your verifier's arbitration behavior between its parametric knowledge and retrieved evidence. The research highlights that this arbitration is highly unreliable and model-dependent, directly impacting factual reliability. You should utilize diagnostic tools like \textsc{PAVE} to assess prior-context discrepancy and consider implementing lightweight test-time arbitration methods, such as the proposed JSD-based approach, to enhance the factual accuracy of your RAG systems without costly model retraining.
Key insights
LLM verifiers in RAG-based fact-checking show unreliable arbitration between parametric priors and contextual evidence, necessitating better evaluation and methods.
Principles
- LLM verifiers have pre-evidence tendencies.
- Prior-context arbitration is model-dependent.
- Factual reliability can be improved post-hoc.
Method
A JSD-based test-time arbitration method improves factual reliability in RAG-based fact-checking. It operates without modifying the underlying LLM, achieving competitive performance across diverse LLM families.
In practice
- Evaluate LLM verifiers with \textsc{PAVE}.
- Select verifiers based on arbitration reliability.
- Apply JSD-based arbitration for RAG.
Topics
- LLM Arbitration
- RAG Fact-Checking
- Verifier Evaluation
- Epistemic States
- JSD Arbitration
- Factual Reliability
Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.