GRACE introduces a framework that deconstructs large language model (LLM) responses into atomic claims and grounds them against trusted knowledge priors within a weighted bipartite graph. Edge weights encode the closeness of each claim to the priors, allowing for classification as Grounded, Refuted, or Boundary.
This classification helps identify hallucinations, as well as novel or contested claims at the frontier of the model's knowledge. The framework uses a Return on Attention (RoA) objective to prioritize claims for expert review based on their uncertainty and verification cost.
Claims verified by experts are used as new evidence anchors, creating an iterative knowledge expansion process. The approach aims to mitigate hallucinations at the system level by combining graph-structured representations with expert-in-the-loop verification.
Evaluation across multiple models and datasets shows that GRACE's knowledge base improves retrieval reliability and efficiently selects boundary knowledge for verification. This method offers a systematic way to enhance model grounding and knowledge accuracy.
Source: https://arxiv.org/abs/2609.04442
