From Feasibility to Desirability: Plan, Learn, Adapt (PLA) Framework for Personalized On-Device Itinerary Generation

2026-07-17 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

The Plan, Learn, Adapt (PLA) framework is a three-stage solution for personalized on-device itinerary generation, addressing the challenge of balancing combinatorial feasibility with subjective traveler preferences under mobile resource constraints. The Plan stage employs a heterogeneous ensemble of lightweight planners to generate diverse, feasible itinerary candidates. Subsequently, the Learn stage utilizes 2,519 pairwise human comparisons across over 100 U.S. cities to fit a compact Bradley-Terry reward model, capturing emergent schedule properties like pacing and geographic coherence. Finally, the Adapt stage refines itineraries locally, preserving feasibility within a device-aware compute budget. PLA achieves a 67.8% win rate, 11.2 percentage points above the best single planner, with 100% feasibility, while frontier LLMs like GPT-5 achieve 0%. Deployed in FlyEnJoy, PLA increased itinerary completion rates by 91% with 109.9 ms on-device latency.

Key takeaway

For AI Engineers developing personalized planning systems, particularly for on-device deployment, you should prioritize a multi-stage approach that explicitly separates feasibility from desirability. Implement an ensemble of lightweight planners to ensure initial feasible candidates, then use human preference data to train a compact reward model for subjective quality. Crucially, your refinement steps must preserve feasibility by construction, as demonstrated by PLA's 100% feasibility and 91% completion rate increase, significantly outperforming LLMs in constrained environments.

Key insights

A multi-stage framework effectively balances combinatorial feasibility with subjective desirability for on-device itinerary generation.

Principles

• Ensemble diverse planners for structural variety.
• Learn preferences from pairwise human comparisons.
• Ensure feasibility at every refinement step.

Method

Generate diverse feasible plans, fit a Bradley-Terry reward model from human comparisons, then apply feasibility-preserving local refinement within device constraints.

In practice

• Employ lightweight planner ensembles for candidate generation.
• Leverage human preference data for reward model training.
• Prioritize feasibility-preserving local search on-device.

Topics

• Personalized Itinerary Generation
• On-Device AI
• Ensemble Planning
• Bradley-Terry Model
• Feasibility Constraints
• Large Language Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Embark Now: User Demand Oriented Framework for Multi-day Urban Travel Itinerary Planning — Integrating LLMs for dynamic user preference capture with an enhanced GRASP algorithm optimizes complex multi-day travel planning.
• Hard Rules, Soft Preferences: Bridging Reasoning, Learning, and Optimization for Personalized Packing Checklist Generation — Combining symbolic reasoning, preference learning, and optimization effectively generates personalized, constrained checklists, outperforming LLMs.
• Reward-Aware Trajectory Shaping for Few-step Visual Generation — Preference alignment awareness in few-step generation can enable student models to surpass teacher performance.
• Top 100 Claude Certified Architect — Professional Questions and Answers — Choosing the optimal LLM architecture depends on task predictability and the need for dynamic adaptation.
• Sparse Personalized Text Generation with Multi-Trajectory Reasoning — PaT enhances cold-start LLM personalization by integrating style and topic cues via iterative, differential-reward-based reasoning.
• Convergence of empirical subgradients for optimal transport-based objectives — Empirical subgradients of optimal transport objectives converge to population subdifferentials, ensuring consistent optimization.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.