Research map / Vision-language and multimodal / Multimodal LLMs
UrbanLLaVA: A Multi-modal Large Language Model for Urban Intelligence
Jie Feng, Shengyuan Wang, Tianhui Liu, Yanxin Xi, Yong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune VILA1.5-8B with UData and UTrain on urban data | VILA1.5-8B · UBench Beijing Geo+Traj | accuracy: 0.0589 -> 0.2800 | Improved |
| UrbanLLaVA versus GPT4o on GeoQA in London and New York | UrbanLLaVA · UBench London/New York GeoQA | Avg. accuracy: -0.73% (London), -7.37% (New York) vs best baseline | Hurt |
| use three-stage training TA->K->Mix instead of default mixed tuning (ablation) | UrbanLLaVA · UBench Beijing | accuracy | Improved |
| order K->TA versus TA->K in two-stage training (ablation) | UrbanLLaVA · UBench Beijing | accuracy | Mixed |
| add mixture learning stage after TA->K (ablation) | UrbanLLaVA · UBench Beijing | accuracy | Improved |
| vary training order of modalities and trained components (ablation) | UrbanLLaVA · UBench | accuracy | No clear effect |
| vary learning rate (ablation) | UrbanLLaVA · UBench | training stability and accuracy | Improved |
| include street view (STV) data in instruction data (ablation) | UrbanLLaVA (one-stage) · UBench Beijing STV-Address | accuracy: 0.4300 -> 0.8550 | Improved |
| include text trajectory and navigation data (ablation) | UrbanLLaVA (one-stage) · UBench Beijing Navigation | success rate: 0.0000 -> 0.3600 | Improved |
| include multi-satellite-image data (ablation) | UrbanLLaVA (one-stage) · UBench Beijing CameraLoc | accuracy: 0.2600 -> 0.5100 | Improved |
| include CityQA textual urban geography data (ablation) | UrbanLLaVA (one-stage) · UBench Beijing | accuracy: GeoQA 0.5409 -> 0.5741 | Mixed |
| train on Beijing only and test in London and New York | UrbanLLaVA · UBench London/New York | performance | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VILA1.5-8B — VILA: Learning Image Aesthetics From User Comments With Vision-Language Pretraining (CVPR 2023) · better on 3 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.89)
- UrbanKGent: A Unified Large Language Model Agent Framework for Urban Knowledge Graph Construction (NeurIPS 2024 · similarity 0.88)
- UrbanFeel:A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.88)
- SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence (ICLR 2026 · similarity 0.87)
- Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.87)
- UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
