Research paper on multi-hop question answering systems reveals that failures can be attributed to retrieval or extraction failures, with extraction failures being a significant bottleneck. This distinction is crucial for improving the performance of AI agents.