Revisiting General Map Search via Generative Point-of-Interest Retrieval
A new generative framework replaces keyword matching with AI reasoning over 10M+ POIs.
A team of researchers from Tencent has introduced GenPOI, a novel generative framework for Point-of-Interest (POI) retrieval on maps. Published on arXiv (2605.03397), the paper addresses the fundamental limitations of traditional map search that relies on surface-level semantic matching. GenPOI leverages large language models (LLMs) to unify heterogeneous search contexts (e.g., user location, query phrasing, time) and POI data into structured sequences, enabling the model to reason about complex, underspecified queries that would stump conventional systems.
The framework introduces two key innovations: a Geo-Semantic POI Tokenization that encodes both geographic coordinates and semantic tags into compact tokens, and a proximity-aware constrained generation strategy that restricts the LLM's output space to ensure retrieved POIs are spatially relevant and valid. Experiments on Tencent Map's industrial-scale dataset—featuring over 10 million POIs—show GenPOI significantly outperforms existing retrieval baselines, particularly for ambiguous queries that require deep contextual reasoning (e.g., "something good for dinner near the park"). This approach reimagines map search as a generative task rather than a lookup problem.
- GenPOI uses LLMs to convert map search into a generative task, outperforming traditional retrieval on ambiguous queries.
- It introduces Geo-Semantic POI Tokenization to encode both location and semantic info into compact token sequences.
- Tested on Tencent Map's real-world dataset of 10M+ POIs, achieving superior performance on context-dependent searches.
Why It Matters
GenPOI makes map search smarter, handling vague or personalized queries that current systems fail to answer.