Xiufeng Liu, 杨艳艳
Navigating multidimensional data cubes remains challenging for analysts who must manually explore vast spaces of possible views through drill-down, roll-up, slice, and dice operations, often missing important patterns or getting lost in uninformative details. While large language models (LLMs) excel at understanding natural language and coordinating tools, they are unreliable for end-to-end data analysis due to hallucination risks in numerical computations. We introduce NavLLM , a system that addresses the next-view recommendation problem by leveraging LLMs as preference models rather than numeric engines. NavLLM formalizes cube navigation as a graph traversal problem and employs a hybrid utility function combining data-driven interestingness computed by a conventional OLAP engine, LLM-estimated preference scores from conversational context, and diversity measures based on navigation history. Given a current view and user utterances, the system generates candidate views, scores them through weighted combination of these three components, and recommends top-k options with natural language explanations. We evaluate NavLLM on three domain-specific cubes (retail, manufacturing, environmental) through automated experiments with 280 navigation sessions across 20 analysis tasks and 7 methods. Results show NavLLM achieves 71% higher cumulative interestingness than LLM-only baselines ( p = . 002 ) and 92% higher than purely data-driven heuristics ( p < .001), while achieving 72% higher hit rate compared to random exploration. The experiments demonstrate that hybrid data-and-preference scoring discovers substantially more valuable patterns during navigation, validating the core design principle of combining statistical interestingness with conversational context.