AI
LLMs Trained to Recommend Your Next Location With Semantic Awareness
A new framework uses reinforcement learning and topology-aware IDs to help large language models suggest places you might want to visit next.
Researchers have developed Refine-POI, a method for training large language models to recommend the next place a person might want to visit. The approach tackles two stubborn problems that plague existing recommendation systems: the inability to maintain semantic relationships between locations when encoding them, and the tendency of supervised models to get stuck on single predictions rather than considering multiple ranked options.
The Core Challenge#
Location recommendation systems typically convert places into numerical IDs so models can process them, but standard approaches don't preserve the underlying meaning of those locations. Two semantically similar venues might end up with ID values far apart, breaking the model's ability to reason about proximity and relationships. Additionally, training methods that rely on matching a single correct answer leave no room for models to generate ranked lists or explain their reasoning—limiting their practical usefulness.
The Technical Solution#
The team introduced two key innovations. First, they use a hierarchical self-organizing map (SOM) to assign semantic IDs to locations. This technique ensures that when venues share similar characteristics in a learned vector space, their assigned ID numbers stay close together, preserving semantic structure.
Second, they employ policy-gradient reinforcement learning to optimize the model's ability to generate ranked recommendation lists rather than forcing it to pick one answer. This approach treats recommendation as a sequential decision-making problem, freeing the model from strict label matching while maintaining accuracy.
Results and Impact#
Testing on three real-world datasets showed that Refine-POI outperformed existing state-of-the-art methods. The framework combines the reasoning and explanation capabilities that LLMs bring to the table with the precise representational fidelity needed for reliable location suggestions, making it useful for travel apps, city navigation, and similar applications where users benefit from understanding why a place is recommended.
