Generative Spatiotemporal Intent Sequence Recommendation via Implicit Reasoning in Amap
The GPlan framework generates logically coherent and physically executable intent sequences for applications like Amap by internalizing Large Language Model reasoning into lightweight models. It uses Progressive Implicit CoT Distillation to handle low-latency generation and Spatiotemporal Counterfactual DPO to ensure plans are sensitive to real-world location and time constraints, leading to improved sequence coherence.