Generating personalized trip itineraries is a complex planning task, involving a tension between hard combinatorial feasibility and soft latent desirability. Classical optimization enforces constraints but fails to capture subjective traveler preferences. Learning-based approaches can model preferences but cannot guarantee feasibility. Mobile deployment adds additional resource constraints on both. To address this, we propose the Plan, Learn, Adapt (PLA) framework for personalized on-device itinerary generation.
In the Plan stage, a heterogeneous ensemble of lightweight planners is built to produce structurally diverse feasible candidates. The Learn stage fits a compact Bradley-Terry reward model from pairwise itinerary comparisons, capturing emergent schedule properties such as pacing, geographic coherence, and day balance that per-POI signals miss. Finally, the Adapt stage applies feasibility-preserving local refinement within a device-aware compute budget, ensuring every intermediate state is feasible by construction.
In tests involving 2,519 pairwise human comparisons across over 100 U.S. cities, the reward-guided ensemble achieved a 67.8% win rate, 11.2 percentage points above the best single planner, with 100% feasibility. Three frontier LLMs, GPT-5, Claude Opus 4.5, and Gemini 3 Pro, achieved 0% feasibility under the same constraints. The reward model generalizes well across held-out cities, with a mean leave-one-city-out accuracy of 67.6%. In production deployment within FlyEnJoy, the PLA framework increased itinerary completion rates by 91%, with an average on-device latency of 109.9 ms.
Blogger's Review: The PLA framework effectively addresses the feasibility and desirability challenge in personalized itinerary generation by integrating traditional optimization with learning models. Its performance in practical applications presents new insights and directions for future personalized services.