Diagnosing LLM Arbitration Behavior over Pre-evidence Epistemic States in RAG-based Fact-Checking

2026-05-31

AI Summary

Current evaluation frameworks struggle to characterize the arbitration behavior of large language models (LLMs) in retrieval-augmented generation (RAG)-based fact verification, particularly when conflicts arise between parametric priors and retrieved evidence. This work proposes PAVE, a diagnostic platform that introduces the first prior-aware verifier evaluation framework. It categorizes models into four cognitive states based on the correctness and confidence of their priors and quantifies their arbitration strategies using Jensen-Shannon divergence (JSD). Experiments across seven mainstream LLMs reveal that existing verifiers exhibit highly model-dependent and unreliable arbitration behavior. The proposed lightweight JSD-based arbitration method, which requires no model modification, significantly improves fact-checking accuracy and demonstrates robust performance across diverse model families.

Abstract

In RAG-based fact-checking, LLMs are increasingly used as verifiers to check given claims against retrieved evidence. Their parametric knowledge can induce pre-evidence tendencies that may conflict with the retrieved context, yet existing evaluation frameworks do not characterize such prior-context discrepancy or measure how verifiers arbitrate between parametric and contextual signals. We introduce \textsc{PAVE} (\emph{Prior-Aware Verifier Evaluation}), a diagnostic testbed that stratifies an LLM verifier into four epistemic states based on the correctness and confidence of its pre-evidence prior and evaluates its arbitration behavior on this new benchmark, i.e., whether it persists in correct prior under misleading evidence, and whether it corrects wrong prior when accurate evidence is provided. Experiments across seven LLMs reveal unreliable and highly model-dependent prior-context arbitration, highlighting the importance of verifier selection for real-world RAG-based fact-checking applications. Based on these findings, we propose a lightweight JSD-based test-time arbitration method that improves factual reliability without modifying the underlying model, achieving competitive performance across diverse LLM families.

Problem

• LLM arbitration
• pre-evidence prior
• RAG-based fact-checking
• epistemic states
• prior-context discrepancy

Innovation

• Prior-Aware Verification
• Epistemic States
• RAG-based Fact-Checking
• Model Arbitration
• JSD-based Calibration