Why AI Still Struggles to Build Travel Itineraries: Map Reasoning Remains a Weak Spot

Large language models can describe destinations and pull travel data, but they still struggle with a basic planning task: figuring out whether an itinerary is physically possible. A July study from researchers including teams at the Chinese University of Hong Kong found that in the TREK benchmark, most AI agents produced schedules that could not actually be completed within the available time, even when they were allowed to query travel-time tools. The top-performing model, GPT-5.6, generated fully feasible plans in only 46.2% of solvable tasks. TREK includes 800 multi-constraint travel-planning tasks built on a database spanning 375 cities and roughly 210,000 records covering flights, hotels, attractions, and rental cars. The benchmark scores outputs against explicit rules rather than AI judges, and the researchers also provide human-verified gold answers. Earlier work pointed in the same direction: TravelPlanner reported that GPT-4 fully satisfied all constraints in just 0.6% of cases, while MapEval found that no tested model exceeded 67% accuracy on map reasoning questions. Researchers and companies are now pursuing three main fixes: connecting models to live map services, pairing language models with mathematical solvers, and training models directly on spatial reasoning tasks. Of those, only the first approach has already reached commercial products for general users.