• Propose a novel LLM-based framework for multi-step location prediction. • Develop an LLM-driven semantic distillation framework to extract behavioral patterns from raw trajectories. • Develop a TAM-MoE architecture to dynamically fuse heterogeneous features. • Validate our framework using two publicly available mobility datasets. • Outperform strong baselines in both per-step and average prediction accuracy. Multi-step human mobility prediction serves as a foundation for bundled service applications such as coordinated tourism planning and dynamic multimodal routing. Existing research mainly focuses on next location prediction, which limits the integration of advanced recommendation strategies and downstream tasks. To bridge this gap, we propose a novel multi-step location prediction framework based on reflection-enhanced distillation and task-aware modular mixture-of-experts (LLM-REDTAM), built upon large language models (LLMs). Our approach integrates LLM reasoning into the predictive framework by constructing semantically rich representations through prompt-based extraction and reflection-driven distillation, which are then fused via a task-aware expert module. We first design two domain-specific prompt templates, converting user check-in sequences into structured textual inputs to extract meaningful semantic representations from raw mobility data and enabling semantic reasoning via a general-purpose LLM. To strategically enhance semantic quality, we introduce a reflection-enhanced distillation approach that leverages iterative self-refinement to enrich LLM-generated insights. This strategy generates a targeted distillation dataset, enabling efficient transfer of comprehensive semantic knowledge from a general-purpose teacher LLM into a specialized, lightweight student model tailored for mobility prediction. In this way, we effectively bridge the gap between the intrinsic complexity of raw mobility data and the generalized knowledge patterns essential for robust predictions. In parallel, we employ a sequential neural model and user embedding to comprehensively capture both immediate spatiotemporal dynamics and individual long-term mobility behavior. To integrate these heterogeneous features, we introduce a task-aware modular mixture-of-experts (TAM-MoE) layer that dynamically fuses semantic features from the LLM, spatiotemporal representations from the sequence model, and personalized user embeddings, flexibly adapting to varying prediction contexts. Experiments on two real-world datasets demonstrate that our method consistently outperforms strong baselines on multi-step prediction tasks, validating its effectiveness for real-world mobility-aware service planning.
Chi et al. (Thu,) studied this question.