TrajVisAgent: Automating Trajectory Visualization from Natural Language Queries via Collaborative Agent Workflow

ACL ARR 2026 January Submission7285 Authors

Keywords: Text-to-TrajVis; Text-to-TrajVis Benchmark; Multi-Agent Systems; Natural Language Processing; Large Language Models

Abstract: The Natural Language to Trajectory Visualization (NL2TrajVis) task aims to automatically translate Natural Language Queries (NLQs) into trajectory data visualizations, thereby enabling natural language interaction within trajectory visualization systems. To overcome the limitations of existing benchmarks in cross-domain coverage and visual aesthetic modeling, we propose TrajVL 2.0, a novel benchmark dataset designed for NL2TrajVis. TrajVL 2.0 covers five representative application domains, comprises 4,552 high-quality samples and explicitly incorporates users' visual aesthetic preferences. Building on this dataset, we further propose TrajVisAgent, a multi-agent collaborative framework tailored for NL2TrajVis. TrajVisAgent comprises a TVL Agent, a Code Agent, and a Visual Agent, which collaboratively handle TVL generation and aesthetic attribute extraction, executable code synthesis with self-repair mechanisms, and iterative optimization guided by visual feedback. This framework enables end-to-end automation, spanning from natural language understanding to trajectory visualization generation and visualization quality refinement. We conduct a systematic evaluation against multiple existing methods on TrajVL 2.0. Experimental results demonstrate that TrajVisAgent consistently outperforms all baseline methods, achieving state-of-the-art performance. Ablation studies further validate the effectiveness of each individual agent as well as their collaborative design.

Paper Type: Long

Research Area: Low-resource Methods for NLP

Research Area Keywords: Data Augmentation

Contribution Types: Data resources

Languages Studied: English

Submission Number: 7285