MODE-RAG: Manifold Outlier Diagnosis and Energy-based Retrieval-Augmented Generation Evaluation

The article introduces MODE-RAG, a Multi-Agent system designed to address hallucinations and logical fabrications in Multimodal Retrieval-Augmented Generation (M-RAG) models. It leverages Variational Free Energy (VFE) and internal attention states to manage interventions dynamically, utilizing Monte Carlo Tree Search (MCTS) for causal reasoning and logit perturbations to mitigate sycophancy. The proposed system, evaluated against the ModeVent benchmark, demonstrates a significant reduction in hallucination rates, enhancing the reliability of M-RAG implementations for practitioners.