New AI Framework Creates Personalized On-Device Itineraries

Generating personalized trip itineraries presents a dual challenge: ensuring the itinerary is feasible (e.g., respecting travel times, opening hours) while also being desirable to the user (e.g., preferred pacing, activity types). Traditional optimization methods can guarantee feasibility but often miss subjective preferences, whereas learning-based approaches capture preferences but struggle with strict constraints. This new Plan, Learn, Adapt (PLA) framework addresses this by integrating three stages, optimized for on-device mobile deployment. The "Plan" stage utilizes an ensemble of lightweight planners to generate a diverse set of structurally feasible itinerary candidates. This ensures that all initial options are practical. The "Learn" stage then employs a compact Bradley-Terry reward model, trained on human comparisons, to capture nuanced user preferences regarding schedule properties like pacing and geographic coherence, which are often missed by simpler point-of-interest signals. Finally, the "Adapt" stage applies feasibility-preserving local refinements to the chosen itinerary within a device's computational limits. This iterative refinement ensures that the itinerary remains feasible at every step while adapting to user preferences. Deployed in a travel app, PLA dramatically increased itinerary completion rates and maintained low on-device latency, outperforming even frontier large language models in feasibility.