From Feasibility to Desirability: Plan, Learn, Adapt (PLA) Framework for Personalized On-Device Itinerary Generation
The PLA framework addresses the conflict between combinatorial feasibility and subjective desirability in itinerary planning through a three-stage process: Plan, Learn, and Adapt. The 'Plan' stage uses a heterogeneous ensemble of lightweight planners to generate structurally diverse yet strictly feasible candidates. The 'Learn' stage employs a compact Bradley-Terry reward model trained on pairwise human comparisons to capture emergent schedule properties like pacing and geographic coherence. The
Analysis
TL;DR
- The PLA framework addresses the conflict between combinatorial feasibility and subjective desirability in itinerary planning through a three-stage process: Plan, Learn, and Adapt.
- The 'Plan' stage uses a heterogeneous ensemble of lightweight planners to generate structurally diverse yet strictly feasible candidates.
- The 'Learn' stage employs a compact Bradley-Terry reward model trained on pairwise human comparisons to capture emergent schedule properties like pacing and geographic coherence.
- The 'Adapt' stage performs feasibility-preserving local refinement within strict on-device compute budgets, ensuring every intermediate state remains valid.
- In production deployment, PLA achieved a 67.8% win rate against single planners, 100% feasibility (compared to 0% for frontier LLMs), and 109.9 ms average on-device latency.
Why It Matters
This research demonstrates a viable path for deploying complex, constraint-heavy generative AI applications on edge devices by decoupling feasibility enforcement from preference modeling. It highlights the limitations of current frontier LLMs in handling hard combinatorial constraints without specialized architectural interventions, offering a blueprint for hybrid systems that combine classical optimization with learned rewards. For practitioners, it provides a concrete methodology for improving user engagement metrics in travel and logistics apps through efficient, on-device personalization.
Technical Details
- Architecture: A three-stage pipeline consisting of Plan (ensemble of lightweight planners), Learn (Bradley-Terry reward model), and Adapt (feasibility-preserving local refinement).
- Reward Modeling: Utilizes pairwise comparisons to train a compact model that captures latent preferences such as pacing, geographic coherence, and day balance, which point-of-interest (POI) signals alone miss.
- Benchmarking: Evaluated on 2,519 pairwise human comparisons across over 100 U.S. cities, showing a 67.6% mean leave-one-city-out accuracy for generalization.
- Performance Metrics: Achieved a 67.8% win rate (11.2 percentage points higher than the best single planner) with 100% feasibility, whereas GPT-5, Claude Opus 4.5, and Gemini 3 Pro achieved 0% feasibility under identical constraints.
- Deployment: Integrated into the FlyEnJoy app, resulting in a 91% increase in itinerary completion rates and an average inference latency of 109.9 ms on mobile devices.
Industry Insight
- Hybrid AI is Essential for Constraint Satisfaction: Pure LLM approaches struggle with hard constraints; integrating classical planning or rule-based layers with learned reward models is necessary for reliable real-world deployment in logistics and planning domains.
- Edge Deployment Enables Privacy and Speed: Moving personalized planning logic to on-device hardware reduces latency and preserves user data privacy, offering a competitive advantage in consumer-facing AI applications.
- Reward Models Capture Nuanced Preferences: Simple scoring mechanisms often fail to capture complex user satisfaction drivers like "pacing" or "coherence"; pairwise comparison-based reward modeling offers a more robust way to align AI outputs with human subjective experience.
Disclaimer: The above content is generated by AI and is for reference only.