ARTFEED — Contemporary Art Intelligence

TrajPrism Benchmark Unifies Language and Trajectory Understanding

ai-technology · 2026-05-12

The introduction of a new benchmark, TrajPrism, aims to connect natural language descriptions with urban trajectory data. This benchmark integrates three distinct tasks: generating trajectories based on instructions, retrieving semantic trajectories driven by language, and captioning trajectories. It combines actual urban trajectories with language annotations that have been filtered by judges, following a four-step methodology. An assessment protocol evaluates trajectory fidelity, the quality of retrieval, and the grounding in language. This research bridges the divide between geometry-focused trajectory modeling and language-oriented mobility benchmarks that prioritize route planning over precise text-route matching. The findings are available on arXiv with the identifier 2605.10782.

Key facts

  • TrajPrism is a multi-task benchmark for language-trajectory alignment.
  • It unifies instruction-conditioned trajectory generation, semantic retrieval, and captioning.
  • The benchmark uses real urban trajectories paired with judge-filtered language annotations.
  • Annotations are generated under a four-step process.
  • Evaluation protocol measures trajectory fidelity, retrieval quality, and language groundedness.
  • Prior work rarely evaluates trajectory and language modalities together on the same real-world data.
  • The research is published on arXiv with identifier 2605.10782.
  • The work aims to improve fine-grained alignment between text and underlying routes.

Entities

Institutions

  • arXiv

Sources