So essentially, WiM consistently outperforms baseline methods (LLM and RAG) in multi-hop reasoning and aggregation tasks.