AgenticRAG-FP Benchmarks Causal Failure Attribution in Multi-Hop RAG

AgenticRAG-FP is an interventional benchmark designed to diagnose where errors occur in multi-hop retrieval-augmented generation. Testing on Claude Haiku 4.5 shows that while coverage-based diagnosis identifies first-hop errors with 0.91 accuracy, it fails to identify errors occurring in later hops.