TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding
Researchers have introduced TrajPrism, a new multi-task benchmark designed to bridge the gap between geometric trajectory modeling and natural language descriptions in urban mobility. Unlike prior works that often treat these modalities separately, TrajPrism evaluates the alignment between spatial trajectories and textual travel intents, constraints, and preferences. The benchmark encompasses three core tasks: instruction-conditioned trajectory generation, language-driven semantic trajectory retrieval, and trajectory captioning. Constructed using 300,000 real-world trajectories from Porto, San Francisco, and Beijing, the dataset yields 2.1 million task instances based on a four-dimensional travel-intent taxonomy. To demonstrate the benchmark's utility, the authors developed proof-of-concept models named TrajAnchor, TrajFuse, and TrajRap. Results indicate that traditional geometry-only baselines perform poorly when language is integrated into the input-output interface, highlighting the necessity for language-grounded approaches. The project includes a reproducible annotation pipeline and code, designed to be portable across different cities with compatible map resources, aiming to advance research in verifiable text-route alignment.
Wire timeline
TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding
Researchers have introduced TrajPrism, a new multi-task benchmark designed to bridge the gap between geometric trajectory modeling and natural language descriptions in urban mobility. Unlike prior works that often treat these modalities separately, TrajPrism evaluates the alignment between spatial trajectories and textual travel intents, constraints, and preferences. The benchmark encompasses three core tasks: instruction-conditioned trajectory generation, language-driven semantic trajectory retrieval, and trajectory captioning. Constructed using 300,000 real-world trajectories from Porto, San Francisco, and Beijing, the dataset yields 2.1 million task instances based on a four-dimensional travel-intent taxonomy. To demonstrate the benchmark's utility, the authors developed proof-of-concept models named TrajAnchor, TrajFuse, and TrajRap. Results indicate that traditional geometry-only baselines perform poorly when language is integrated into the input-output interface, highlighting the necessity for language-grounded approaches. The project includes a reproducible annotation pipeline and code, designed to be portable across different cities with compatible map resources, aiming to advance research in verifiable text-route alignment.
cs.AI updates on arXiv.org