New Framework Generates Personalized On-Device Travel Itineraries

This paper presents the Plan, Learn, Adapt (PLA) framework, a three-stage approach for generating personalized trip itineraries directly on mobile devices. The core challenge lies in simultaneously satisfying hard combinatorial constraints (feasibility) and capturing subjective traveler preferences (desirability), all while adhering to mobile resource limitations. Traditional optimization methods often miss preferences, while learning-based approaches struggle with feasibility guarantees. The PLA framework addresses this by first using a "Plan" stage, where an ensemble of lightweight planners generates diverse, feasible itinerary candidates. Next, the "Learn" stage fits a compact Bradley-Terry reward model from pairwise human comparisons. This model captures emergent 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 within a device's compute budget, ensuring all intermediate states remain valid. Evaluations across over 100 U.S. cities and 2,519 human comparisons showed the reward-guided ensemble achieved a 67.8% win rate, significantly outperforming single planners, with perfect feasibility. Notably, frontier LLMs achieved 0% feasibility under the same constraints. Deployed in the FlyEnJoy app, PLA boosted itinerary completion rates by 91% with minimal on-device latency.