Personalized Itinerary Planning with LLM-Guided Interaction

Introduction: Personalized itinerary planning requires systems capable of interpreting diverse user preferences while minimizing cognitive effort. Traditional tools often rely on rigid form-based inputs or unstructured conversational agents, which limit usability. This study introduces a novel hybrid interface guided by Large Language Models (LLMs) that combines conversational flexibility with structured controls, enhanced through implicit behavioral feedback, to improve travel planning efficiency. Methods: A system integrating a fine-tuned LLaMA-3 model with a hybrid interface combining conversational and form-based inputs was developed. Implicit feedback (dwell time, clicks) was incorporated via a weighted scoring algorithm to refine recommendations in real time. A user study involving 50 participants was conducted to compare three interfaces: conversational, form-based, and hybrid. Standardized metrics were used, including usability (SUS), cognitive load (NASA-TLX), efficiency (the completion time and the number of iterations), and satisfaction. Results: The hybrid interface achieved a significantly higher usability score than conversational or form-based interfaces (86.1 ± 5.9), lowest cognitive load (28.9 ± 4.7), and fastest itinerary completion time (11.8 ± 1.6 minutes). Recommendation accuracy reached 92%, supported by a 28% reduction in planning iterations due to the integration of implicit feedback. Satisfaction ratings were highest for the hybrid interface (4.6 ± 0.5 on a 5-point scale), with 92% of participants indicating they were satisfied or very satisfied. Discussion: Findings confirm that hybrid interfaces strike a balance between conversational flexibility and structured clarity, substantially improving the user experience compared to traditional approaches. Implicit feedback mechanisms reduce user effort and enhance the relevance of recommendations, demonstrating the advantage of adaptive interaction. Limitations include occasional LLM misinterpretations and minor latency (mean 1.2 seconds per query). Broader benchmarking against traditional recommender baselines suggests clear efficiency gains, though further generalization across domains is warranted. Findings should be interpreted cautiously, as the study was limited to 10 destinations and had a modest participant pool (N = 50). Future work should investigate multimodal inputs (such as voice and image) and conduct longitudinal studies to assess sustained engagement. Conclusion: LLM-guided hybrid interfaces provide a scalable and effective solution for personalized itinerary planning, significantly optimizing usability, accuracy, and efficiency. By integrating implicit behavioral signals with adaptive human–computer interaction, this work advances both intelligent recommendation systems and HCI research, with potential applicability to related domains beyond travel.