Amap's GPlan distills LLM reasoning for intent sequence recommendations

Real-world user behavior isn't a series of isolated actions—it's a flow of intents tied to time and space. Think of a driver leaving work: they might need to pick up groceries, drop off a package, and grab coffee, all in a logical order that respects traffic and business hours. The task of Generative Spatiotemporal Intent Sequence Recommendation (GSISR) aims to produce such multi-step plans. While large language models (LLMs) excel at reasoning, their direct use in production (e.g., in Amap, China's leading navigation app) is hampered by high inference latency and a tendency to generate plans that are either context-mismatched or physically impossible.

To bridge this gap, the authors introduce GPlan, a lightweight framework that internalizes LLM reasoning. The first component, Progressive Implicit CoT Distillation, compresses the explicit chain-of-thought reasoning from a teacher LLM into a small set of latent tokens. This allows a smaller student model to inherit complex planning logic without generating long reasoning text—dramatically cutting latency. The second component, Spatiotemporal Counterfactual DPO, fine-tunes the model using counterfactual context-plan pairs (e.g., swapping a user's location or time) to teach sensitivity to real-world constraints. The result: a model that produces logically coherent, physically executable intent sequences. Offline experiments and online A/B testing on Amap confirmed gains in sequence coherence and context responsiveness.