AnalysisAI ModelsJuly 2, 2026
New papers tackle RAG hallucination detection
Papers propose span-level hallucination detection (Beyond Document Grounding, CORTEX), budget-constrained multi-hop RAG (What Survives Into Context), confidence-based decoding (Dual-Confidence), and factual error repair (Diagnosing and Repairing). CORTEX performs token-level detection via comparative internal representations.