Amap's Gwhere boosts POI recommendations 6% with generative LLM

Researchers at Amap (Alibaba's mapping and navigation platform) have introduced Gwhere, an end-to-end generative retrieval framework for next Point-of-Interest (POI) recommendation. The system addresses a key challenge: representing complex spatial entities like restaurants or transit stations as compact tokens that an LLM can retrieve. Gwhere first trains a contrastive residual-quantization tokenizer to convert multimodal POI data—textual descriptions, images, geographic coordinates, and collaborative signals—into discriminative semantic identifiers (SIDs). These SIDs are then used to adapt an LLM to mobility scenarios through continued pretraining on enriched spatio-temporal corpora, supervised fine-tuning, and a novel reinforcement learning objective called Exposure-Aware Kahneman-Tversky Optimization (EAKTO). EAKTO aligns the model's predictions with real user behavioral preferences, penalizing unrealistic or overexposed recommendations.

Deployed in Amap's homepage service under high-concurrency and low-latency constraints, Gwhere has been tested in long-term online A/B experiments. Results show a 5.83% improvement in Page Click-Through Rate (P-CTR) and a 6.20% improvement in User Click-Through Rate (U-CTR) compared to the production baseline. The paper is published on arXiv and the implementation is open-sourced. This work demonstrates that generative retrieval via LLMs can scale to industrial-scale recommendation systems, particularly for structured spatial data. The approach overcomes token redundancy and collision problems common in prior generative recommenders, and it integrates heterogeneous signals more effectively than existing methods.