Map applications bridge everyday mobility with local services, creating opportunities to recommend places by uncovering users' implicit intents. However, utility-driven map interactions provide only indirect and incomplete evidence of consumption demands, making it challenging to infer personalized service needs from observed behaviors. We propose MILES-Rec, a cross-domain POI recommendation framework that infers latent service intents from heterogeneous map behaviors and expresses them as prioritized natural language queries. It first learns from progressive reasoning demonstrations that synthesize user profiles and map histories for situation understanding, intent expansion, and query prioritization. We then introduce intent-aligned group relative policy optimization, using subsequent user interactions to reward query relevance and ranked service coverage while penalizing redundancy and excessive response length. A two-level credit assignment mechanism integrates shared list-level advantages with query-specific advantages, coordinating individual query quality with the complementary coverage of the complete query list. We further construct MILES-Bench from real user interaction logs on a large-scale mapping platform, supporting training and evaluation across diverse services under rich and sparse behavioral histories. Extensive experiments against general-purpose LLMs and specialized recommendation baselines demonstrate superior POI retrieval utility and strong service and semantic alignment, with consistent gains from behavioral policy optimization across four backbone scales.

翻译：暂无翻译