AI chatbots can draft a travel plan, but they often fail on the part that matters most in practice: whether the schedule actually fits into a day. Travel time between stops is frequently misjudged, and the result is an itinerary that looks coherent on paper but cannot be completed in real life.
A July study from research teams including the Chinese University of Hong Kong shows that the problem remains unresolved even in newer models. In the TREK benchmark, 15 AI agents were tested on multi-constraint travel-planning tasks. Most still produced schedules that could not be completed within the available time, even though they were allowed to use tools to query travel duration.
TREK contains 800 travel-planning tasks with multiple constraints. Of those, 533 are solvable and 267 were proven unsolvable. The simulated database spans 375 cities and includes about 210,000 records covering flights, hotels, attractions, and rental cars. Agents must query tools for information, including a tool that calculates travel time between two locations, and then produce itineraries that satisfy budget, trip length, opening hours, and traveler preferences that are not explicitly stated.
Scoring is rule-based rather than handled by AI judges. The researchers also provide human-verified reference answers that can achieve full marks, showing that perfect scores are possible.
GPT-5.6 led the ranking, but still cleared fewer than half of solvable tasks
GPT-5.6 achieved full marks on 46.2% of solvable tasks, the best result in the benchmark. The second-place model, Kimi-K2.5, scored 17.1%, while GLM-5 reached 13.1%. Across all 15 agents, the median score was just 6.6%, and the lowest result was 0. Claude and Gemini were not included in this test.
The metric most directly tied to map reasoning asks whether the gap between two same-day activities is shorter than the actual travel time required. Depending on the model, the share of tasks with errors on that metric ranged from 13.3% to 93.8%, the widest spread among all scoring categories. The paper describes this as the dividing line between models that can plan and models that cannot.
For GPT-5.6, the main unresolved issue was meeting traveler needs that were left unstated.
Earlier benchmarks pointed to the same weakness
The finding is consistent with prior work. In TravelPlanner, published in 2024, GPT-4 produced itineraries that fully satisfied all constraints in only 0.6% of cases. In MapEval, published in 2025, researchers used 700 questions to test map reasoning across 54 countries and 180 cities. None of the 30 tested models exceeded 67% accuracy, and all trailed humans by more than 20 percentage points. The weakest areas were distance, direction, and route planning.
Models appear to hold spatial representations, but they do not compute routes on road networks
That does not mean language models know nothing about geography. In the 2023 paper Language Models Represent Space and Time, MIT researchers Wes Gurnee and Max Tegmark analyzed Meta's Llama-2 and found that the model internally contains spatial representations for locations around the world, across the United States, and within New York City. Those representations could be read out with simple linear methods.
The problem, according to a May analysis from map provider HERE, is computation. Language models do not calculate routes over road networks, do not evaluate traffic in real time, and cannot guarantee that geometric constraints between coordinates hold. HERE's explanation is straightforward: language models are built to predict text, while spatial reasoning requires precise computation over geographic data. Real travel time depends on roads, schedules, transfers, and opening hours, and that information changes constantly rather than sitting still in training data.
Three approaches are emerging to close the gap
1. Let the model query map services directly
This is the only path that has already reached commercial deployment. In October 2025, Google introduced Grounding with Google Maps in the Gemini API, allowing models to query addresses, opening hours, and reviews for more than 250 million places. Google's example was travel planning: ask AI to build a one-day itinerary in San Francisco, and the response includes distance, travel time, and opening hours for that day.
Google also offers a hosted Maps Grounding Lite MCP server. AI tools that support the MCP protocol can connect to it to search places, check weather, and calculate routes, although routing currently supports only driving and walking. HERE argues for the same division of labor: let the model interpret the request and orchestrate tools, while the map engine handles routes and distance.
2. Combine language models with mathematical solvers
A NAACL 2025 paper from the MIT Aeronautics and Astronautics group led by Chuchu Fan takes that route. The system first translates travel requests into mathematical constraints, then passes them to a solver that can be proven correct. On TravelPlanner, the method reached a 93.9% success rate. The same paper cites a comparison figure showing that OpenAI's reasoning model o1-preview managed only 10%, even when all required information was provided.
When constraints conflict and no valid itinerary exists, the system can also identify which requirements are incompatible and suggest revisions. It successfully repaired 81.6% and 91.7% of unsatisfiable requests on two datasets.
3. Train models to learn spatial reasoning directly
This line of work remains mostly in research. MapTrace, from researchers at Google and the University of Pennsylvania, uses automatically generated data to create 23,000 path samples across 4,000 maps, training a multimodal model to draw traversable routes on maps. On MapBench, the method improved success rates by as much as 6.4 percentage points.
An ACL 2026 paper from Fudan University and ByteDance Seed tested whether a model could reconstruct a full map from fragmentary descriptions such as "home to school is 1.4 kilometers at a bearing of 135 degrees." The results showed that the model could infer relationships between unseen locations and learn road connectivity from driving trajectories before planning routes.
Niantic Spatial, spun out from Pokémon GO developer Niantic, is training what it calls a large geospatial model on large volumes of geolocated real-world imagery to answer questions such as where a user is, which direction they are facing, and what they are looking at. World Labs, founded by Fei-Fei Li and set to be acquired by AMD, is working on "spatial intelligence" focused on generating and understanding explorable 3D scenes, a different branch from the geographic reasoning needed for itinerary planning.
For now, map lookup is the only option ordinary users can access directly
Among the three approaches, only the first has produced services that general users can actually use today. Even so, the TREK results suggest that giving AI access to travel-time tools does not solve the problem by itself. After an itinerary is generated, checking each leg in a map app remains the safer way to verify whether the plan is feasible.
The article cites TREK (arXiv 2607.26977), TravelPlanner, MapEval, and research or analysis from MIT, Google, the University of Pennsylvania, Fudan University, ByteDance Seed, and HERE.