HawkBench is a novel benchmark for evaluating the adaptive resilience of RAG systems to meet the dynamic and diverse needs of users in real-world information retrieval scenarios. Unlike existing benchmarks that focus on specific task types (primarily factual questions) and diverse knowledge bases, HawkBench systematically categorizes a wide range of question types, including factual and evidence-based questions. It integrates multi-domain corpora across all task types to mitigate corpus bias and provides rigorous annotations for high-quality evaluation. It includes 1,600 high-quality test samples, evenly distributed across domains and task types. We evaluate representative RAG methods to analyze their performance in terms of answer quality and response latency, highlighting the need for dynamic task strategies that integrate decision-making, query interpretation, and overall knowledge understanding to improve RAG generalization.
