TL;DR
GLIDE is a production-ready forecasting framework designed for the highly dynamic world of on-demand food delivery. By combining inductive graph quantization, volume-adaptive Mixture-of-Experts, and LLM-distilled event knowledge, it accurately predicts order volumes across evolving logistics networks while achieving a 15x speedup over standard Transformer architectures.
What Problem Does This Paper Address?
In platforms like Meituan, efficient delivery relies on "Order Pooling"—bundling multiple orders for a single courier. To do this effectively, the system must forecast demand at Scale-Effect Hotspots (SEHs): groups of logistics flows that act as optimization units.
However, traditional spatiotemporal models (like those used for traffic on fixed road networks) fail here for three reasons:
- Dynamic Topology: Logistics graphs are ephemeral. Hotspots emerge and vanish as merchant supply and user preferences shift. Standard models that bind parameters to fixed node IDs cannot handle "unseen" nodes.
- Scale–Volume Heterogeneity: Graph sizes vary by orders of magnitude ($10^2$ to $10^4$ nodes), and demand is "zero-inflated" and "long-tailed." Most areas have low volume, but a few "head" hotspots have massive bursts. Standard regression models tend to ignore these critical bursts and "forecast toward the mean" (zero).
- Event-Driven Disruption: Holidays, sudden rainstorms, or campus calendars cause massive demand shifts that historical data alone cannot predict. Retraining a massive model for every rainstorm is operationally impossible.
Methods
The authors propose GLIDE (Graph-Latent Inductive Dynamics prediction with Event awareness), which consists of three core innovations:
- Inductive Graph Quantization (IGQ): To handle variable graph sizes, GLIDE doesn't operate on physical nodes directly. Instead, it learns a sparse assignment matrix to project any physical subgraph into a fixed-size set of $K$ "Latent Tokens." This acts as a topological bottleneck, filtering noise and decoupling computational cost from the number of nodes.
- Volume-Adaptive Mixture-of-Experts (MoE) Head: To solve the long-tail bias, GLIDE uses a committee of specialized experts. A gating network routes representations to experts trained specifically for different volume regimes (e.g., high-volume bursts vs. sparse low-volume demand).
- Event-Aware Modulation: Rather than retraining the model for events, the authors use an offline LLM (GPT-4o-mini) to distill event-impact reasoning into a static knowledge bank. At inference time, a lightweight modulator applies multiplicative corrections to the frozen backbone's predictions based on this bank.
Datasets
The model was evaluated on six large-scale, real-world datasets from the Meituan OFD platform, covering a full year (July 2025 – July 2026).
- Cities: Data from three distinct cities (A, B, and C) with varying scales.
- Scenarios: Split into "Weekday Routine" (stable dynamics) and "Weekend Lifestyle" (bursty dynamics).
- Heterogeneity: Subgraph sizes ranged from a single node to nearly 6,000 nodes, providing a rigorous test for inductive generalization.
Key Insights and Ideas
- The "Simplicity vs. Complexity" Paradox: The authors found that simple linear baselines often outperform complex Transformers on noisy logistics data. GLIDE overcomes this by using its latent token space as a regularization bottleneck, allowing the Transformer to focus on macro-level dynamics rather than node-level noise.
- Zero-Shot Generalization: Because GLIDE learns "Logistics Archetypes" (e.g., "Residential morning peak") in a latent space rather than city-specific node identities, it can be trained in one city and deployed in another with almost no loss in accuracy. In some cases, the zero-shot model even outperformed models trained specifically on the target city due to the richer diversity of patterns in the source city.
- LLM as an Offline Teacher: A major practical insight is moving the LLM out of the inference path. By using the LLM to build a "Knowledge Bank" offline, they get the reasoning power of a foundation model without the latency or cost of calling an API during real-time dispatch.
Key Results
- Accuracy: GLIDE outperformed SOTA baselines (including iTransformer, STGCN, and UniFlow) in 10 out of 12 metrics across multiple cities.
- Efficiency: GLIDE achieved a 15x speedup in training time compared to standard Transformers (18 minutes vs. 4.5+ hours) and demonstrated strictly linear scaling with respect to graph size.
- Online Impact: In A/B tests on Meituan's production system, GLIDE-guided dispatch reduced single-order waybills by 32.46 percentage points, significantly increasing the proportion of efficient multi-order bundles.
- Event Handling: The Event-Aware Modulation provided significant gains during disruptions, such as a +7.3% improvement during Labor Day and +1.6% during sustained heatwaves.
Limitations and Future Work
- LLM Sensitivity: The authors noted one failure case (Qixi Festival) where the LLM-generated impact vector overestimated demand shift, leading to a slight performance drop. This suggests the system is sensitive to the quality of the distilled knowledge.
- Fixed Latent Space: While the latent space size ($K=100$) works for current city scales, extremely large mega-cities might eventually require a more dynamic way to scale the latent bottleneck.
- Future Directions: The authors suggest exploring more granular event descriptions and potentially integrating the event knowledge bank more deeply into the latent space modeling rather than as a post-hoc modulation.
Summary updated