Abstract

This disclosure describes route-planning techniques that leverage large reasoning models to predict route traversal costs based on user-generated content (UGC) and data available in digital map applications. Large reinforcement models, which can advantageously be trained using reinforcement learning (RL), can use sparse and hard-to-interpret signals arising from UGC to extend and to customize routing algorithms for best-path or travel-time queries. Per the techniques, large reasoning models steer the routing algorithm to more effectively estimate the optimal route by introducing user-generated data (e.g., reviews, comments, photos, etc.) into the reasoning chains of a large reasoning model; by causing the large reasoning model to propose modifying costs (e.g., adding rewards or negative costs) for parts of the road graph; and by comparing the actual travel time with the estimated travel time to derive a reward signal that causes the model to surface those reasoning chains that were most useful.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS