Researchers have developed TRIAD, a method for automatically generating datasets for evaluating multi-hop question-answering systems. These systems, known as RAG (Reinforced Autoregressive Generators), are used in industry to answer complex questions based on large amounts of text. Existing datasets like HotpotQA only work with Wikipedia-based knowledge and can't be directly applied to proprietary data. TRIAD generates three types of data: question-answer pairs, relevance-labeled context documents, and unanswerable questions. The method is evaluated against established datasets MuSiQue and HotpotQA, showing similar performance trends across different RAG setups. Human validation indicates that the generated questions are suitable for evaluating domain-specific RAG systems. --- Why it matters: This matters to engineers working on multi-hop question-answering systems because it provides a way to automatically generate high-quality datasets for evaluation, which is essential for improving system performance and adaptability to different domains. Source: https://arxiv.org/abs/2608.21558