Multi-Hop Reasoning

Multi-hop reasoning asks whether an AI system can connect several facts or pieces of evidence, such as linking entities in a knowledge graph or combining passages to answer a question. Common approaches include graph traversal, retrieval-augmented generation with query decomposition and adaptive search, and explicit verification of intermediate steps; some work also trains models with synthetic chains or reinforcement learning. Benchmarks such as HotpotQA and MuSiQue test text-based question answering, while newer tasks probe long documents, visual evidence, temporal graphs, and culturally specific rules. Systems are improving, but success remains sensitive to retrieval and fact extraction, and correct final answers do not always show that a model used the intended evidence or can generalize to longer chains.