MetaRAG: Belief-Action Aligned Policy Optimization for Agentic RAG

MetaRAG optimizes agentic RAG by aligning a model's internal belief about evidence sufficiency with its retrieval actions. It utilizes Verify-first Action Generation and Internal Belief Probing during training to prevent the reinforcement of internally consistent but factually incorrect trajectories.