Next point-of-interest (POI) recommendation typically draws on both a user's visit history and the content of candidate locations, such as descriptions and photos. In practice, those text and image modalities are frequently incomplete. The paper, released on arXiv, addresses this by applying large language model (LLM) agents to the recommendation task, using them for preference modeling and evidence-related tasks.

The authors introduce a task-hypergraph framework that learns complementary groups of agents. Rather than treating all agents as interchangeable, the hypergraph captures how different tasks relate to different agents, allowing the model to assign roles that cover missing information. The approach is designed to make next-POI recommendation more robust when modality data is absent.

The available abstract does not include quantitative results, so the practical gains over existing methods are not yet clear from the description alone.