benchmark stringlengths 5 19 | abstract stringlengths 78 236 | paper_title stringlengths 62 120 | publication_year int64 2.02k 2.03k | venue stringlengths 9 40 | research_group stringlengths 44 131 | paper_url stringlengths 32 149 | code_url stringlengths 26 67 β | data_url stringlengths 39 158 β | url_status stringclasses 9
values | domain stringlengths 25 78 | task_type stringlengths 53 156 | model_type stringclasses 3
values | sample_count stringlengths 26 124 | metrics stringlengths 34 214 | notes stringlengths 90 505 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
GeoBench | Evaluates geoscience knowledge understanding and utilization by foundation language models through geoscience QA and reasoning tasks. | K2: A Foundation Language Model for Geoscience Knowledge Understanding and Utilization | 2,024 | WSDM 2024 | Shanghai Jiao Tong University NLP group (Zhouhan Lin and Junxian He), with Waterloo and CAS IGSNRR collaborators | https://dl.acm.org/doi/10.1145/3616855.3635842 | https://github.com/davendw49/k2 | https://huggingface.co/datasets/daven3/geobench | verified_public_urls | geoscience; Earth science | geoscience QA; knowledge reasoning; domain understanding | LLM | paper: >1,500 objective questions + 939 subjective questions; local: 2,653 entries | accuracy; GPTScore; perplexity; human ratings | Included as the geoscience text benchmark associated with K2; GST-adjacent rather than a dedicated spatiotemporal reasoning benchmark. Paper, README, and local files have minor count discrepancies, especially for NPEE multiple-choice and local true/false/QA categories. |
CityBench | Evaluates large language models on diverse urban tasks involving city knowledge, spatial reasoning, urban planning, mobility, and prediction-style questions. | CityBench: Evaluating the Capabilities of Large Language Models for Urban Tasks | 2,025 | KDD 2025 Datasets and Benchmarks | Tsinghua University FIB Lab / BNRist urban computing group (Yong Li), with BJTU collaborators | https://arxiv.org/abs/2406.13945 | https://github.com/tsinghua-fib-lab/CityBench | https://huggingface.co/datasets/Tianhui-Liu/CityBench-CityData | verified_public_urls | urban science; mobility; maps; traffic | urban QA; spatial reasoning; mobility prediction; planning; city knowledge | LLM | Large urban-task benchmark; task-level counts vary across released modules | accuracy; F1; MAE; RMSE; task-specific metrics | Included as a broad urban LLM benchmark; some components are prediction/evaluation tasks rather than pure QA reasoning. |
UrBench | Evaluates multimodal large models in multi-view urban scenarios using images and urban-context questions across perception and reasoning tasks. | UrBench: A Comprehensive Benchmark for Evaluating Large Multimodal Models in Multi-View Urban Scenarios | 2,025 | AAAI 2025 | Shanghai AI Laboratory OpenDataLab and Sun Yat-sen University urban vision group, with SenseTime and Wuhan University collaborators | https://ojs.aaai.org/index.php/AAAI/article/view/33163 | https://github.com/opendatalab/UrBench | https://huggingface.co/datasets/bczhou/UrBench | verified_public_urls | urban science; street-view; maps | urban scene understanding; multi-view reasoning; visual QA; spatial reasoning | MLLM | Multi-view urban benchmark; exact public data count not verified in stable repository | accuracy; task-specific visual QA metrics | Official code repository and Hugging Face dataset are verified; included as a public AAAI benchmark for multi-view urban MLLM evaluation. |
MapEval | Evaluates foundation models on map-based geospatial reasoning across textual, API-based, and visual map interaction settings. | MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models | 2,025 | ICML 2025 | BUET CSE map reasoning group (Mohammed Eunus Ali), with Monash University and QCRI collaborators | https://proceedings.mlr.press/v267/dihan25a.html | https://github.com/MapEval | https://huggingface.co/MapEval/datasets | verified_public_urls | maps; navigation; urban science; geospatial reasoning | map understanding; spatial relationship reasoning; navigation; travel planning; API/tool reasoning | LLM; MLLM | 700 multiple-choice questions across 180 cities and 54 countries | accuracy; multiple-choice accuracy | Core GSTA benchmark for map/tool/visual geospatial reasoning. Project page links the official GitHub organization and Hugging Face data organization. |
GEOBench-VLM | Benchmarks vision-language models on a broad suite of geospatial and remote-sensing tasks using curated image-instruction pairs. | GEOBench-VLM: Benchmarking Vision-Language Models for Geospatial Tasks | 2,025 | ICCV 2025 | MBZUAI vision/geospatial group, with UCL, Linkoping University, IBM Research, ServiceNow, and ANU collaborators | https://openaccess.thecvf.com/content/ICCV2025/html/Danish_GEOBench-VLM_Benchmarking_Vision-Language_Models_for_Geospatial_Tasks_ICCV_2025_paper.html | https://github.com/The-AI-Alliance/GEO-Bench-VLM | https://huggingface.co/datasets/aialliance/GEOBench-VLM | verified_public_urls | remote sensing; geospatial; Earth observation | remote sensing VQA; image understanding; temporal understanding; relationship prediction; grounding; segmentation | MLLM | 10k+ human-verified instructions; 8 categories and 31 subtasks | accuracy; IoU; F1; BLEU; CIDEr; task-specific metrics | Included as a major remote-sensing/geospatial VLM benchmark; more perception-heavy than map-reasoning benchmarks but still directly GST-relevant. |
XLRS-Bench | Evaluates multimodal large models on very-high-resolution remote sensing understanding tasks. | XLRS-Bench: Could Your Multimodal LLMs Understand Extremely Large Ultra-High-Resolution Remote Sensing Imagery? | 2,025 | arXiv 2025 | NUDT remote sensing team with Tsinghua NLP group (Zhiyuan Liu and Maosong Sun), BUPT and Wuhan University collaborators | https://arxiv.org/abs/2503.23771 | https://github.com/AI9Stars/XLRS-Bench | https://huggingface.co/collections/initiacms/xlrs-bench | verified_public_urls | remote sensing; Earth observation | remote sensing VQA; high-resolution image understanding; visual reasoning | MLLM | Benchmark scale not fixed in this registry row; verify against repository release | accuracy; task-specific visual QA metrics | Included as an adjacent remote-sensing MLLM benchmark; official GitHub README links the Hugging Face dataset collection. |
STARK | Benchmarks spatiotemporal reasoning capabilities and failure modes of LLMs and reasoning-oriented language models using structured spatiotemporal tasks. | Benchmarking Spatiotemporal Reasoning in LLMs and Reasoning Models: Capabilities and Challenges | 2,025 | NeurIPS 2025 Datasets and Benchmarks | UCLA ECE / NESL group led by Mani Srivastava | https://proceedings.neurips.cc/paper_files/paper/2025/hash/5ebcbc5e7470ff60057600fe43f73b4c-Abstract-Datasets_and_Benchmarks_Track.html | https://github.com/nesl/STARK_Benchmark | https://huggingface.co/datasets/prquan/STARK_10k; https://huggingface.co/datasets/prquan/STARK_1k | verified_public_urls | spatiotemporal reasoning; mobility; urban science | state estimation; relation reasoning; navigation; ETA; POI prediction; intent prediction | LLM | 14,552 challenges; STARK-L ~14k samples; STARK-S ~1.3k samples | RMSE; RMSPE; trimmed error; code-interpreter relative error change | Core GSTA benchmark for text/structured spatiotemporal reasoning. README and paper differ slightly on task count wording; relation combinations may expand into many implementation-level task IDs. |
TurnBack | Evaluates large language models' geospatial route cognition through reverse-route reasoning tasks. | TurnBack: A Geospatial Route Cognition Benchmark for Large Language Models through Reverse Route | 2,025 | EMNLP 2025 main conference | Huawei Riemann Lab, with Technical University of Munich, Hochschule Karlsruhe, and MCML collaborators | https://aclanthology.org/2025.emnlp-main.440/ | https://github.com/bghjmn32/EMNLP2025_Turnback | https://github.com/bghjmn32/EMNLP2025_Turnback/tree/main/36kroutes | verified_repo_data | navigation; maps; route cognition | route reasoning; reverse navigation; spatial reasoning | LLM | 36k routes across 12 metropolises | accuracy; route-validity metrics; task-specific metrics | Focused LLM navigation/route-cognition benchmark; reproduction is not fully text-only and may require stable OpenStreetMap/Overpass access or prebuilt local graph caches. |
TripCraft | Benchmarks LLMs on spatio-temporally fine-grained travel planning with itinerary constraints and location-time reasoning. | TripCraft: A Benchmark for Spatio-Temporally Fine Grained Travel Planning | 2,025 | ACL 2025 main conference | IIT Bhubaneswar NLP/AI group (Abhik Jana and Shreya Ghosh), with Microsoft India collaborators | https://aclanthology.org/2025.acl-long.834/ | https://github.com/Soumyabrata2003/TripCraft | https://github.com/Soumyabrata2003/TripCraft#-dataset-access | data_access_request_required | travel planning; navigation; mobility; urban science | planning; itinerary generation; spatiotemporal reasoning; constraint satisfaction | LLM | Travel-planning benchmark; dataset/database access appears request-based or repository-mediated | constraint satisfaction; feasibility pass rates; temporal/spatial/ordering/persona continuous metrics; GPT-4o postprocess for plan-to-JSON conversion | Official repo requires request-based access for the dataset and auxiliary databases. The released evaluation scripts consume structured JSONL, while the generation script outputs natural-language itineraries. The README states GPT-4o is used for plan-to-JSON postprocessing, but a complete TripCraft-specific conversion ... |
GeoBenchX | Evaluates large language model agents on GIS-oriented geospatial workflow generation and analysis tasks. | GeoBenchX: Benchmarking LLMs in Agent Solving Multistep Geospatial Tasks | 2,025 | ACM SIGSPATIAL GeoGenAgent 2025 workshop | World Bank and JPMorgan Chase practitioner GeoAI collaboration | https://arxiv.org/abs/2503.18129 | https://github.com/Solirinai/GeoBenchX | https://github.com/Solirinai/GeoBenchX/tree/main/data | verified_repo_data | GIS; geospatial analytics; maps | GIS workflow generation; code generation; geospatial analysis; tool use | LLM | Benchmark/task collection for geospatial code generation; exact count varies by repository version | pass rate; task success; execution accuracy; LLM judge | Included as an agentic GIS benchmark; benchmark data directory is available in the official GitHub repository. |
GeoAnalystBench | Benchmarks AI geospatial analysts on end-to-end GIS analysis, tool use, and geospatial reasoning tasks. | GeoAnalystBench: A GeoAI benchmark for assessing large language models for spatial analysis workflow and code generation | 2,025 | arXiv 2025 | University of Wisconsin-Madison Geospatial Data Science Lab (Song Gao), with Ohio State CSE collaborators | https://arxiv.org/abs/2509.05881 | https://github.com/GeoDS/GeoAnalystBench | https://github.com/GeoDS/GeoAnalystBench/blob/master/dataset/GeoAnalystBench.csv; https://drive.google.com/drive/u/0/folders/1GhgxWkNVh4FTgS1RETgvbstBqx0Q9ezp | verified_public_urls | GIS; geospatial analytics; maps | geospatial analysis; tool use; GIS workflow; code generation; reasoning | LLM | Benchmark task set; exact public count should be verified from repository release | task success; execution accuracy; pass rate; LLM judge | Included as an adjacent geospatial-agent benchmark; official README links both the benchmark CSV and supporting data folder. |
USTBench | Benchmarks and dissects the spatiotemporal reasoning capabilities of LLMs acting as urban agents across understanding, forecasting, planning, and reflection tasks. | USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning Capabilities of LLMs as Urban Agents | 2,026 | ICLR 2026 main conference | HKUST(GZ) AI Thrust / USAIL urban AI group (Hao Liu) | https://openreview.net/forum?id=ETzBStUFJy | https://github.com/usail-hkust/USTBench | https://huggingface.co/datasets/Haruto2099/USTBench-Dataset | verified_public_urls | urban science; spatiotemporal reasoning; traffic; mobility | spatiotemporal reasoning; forecasting; planning; reflection; urban QA | LLM | 62,466 structured QA pairs; 9 downstream tasks | accuracy; MAE; RMSE; MAPE; task-specific metrics | Core urban-agent benchmark; official GitHub README links the Hugging Face dataset release. |
FRIEDA | Evaluates vision-language models on multi-step cartographic reasoning over maps, including topological, metric, and directional reasoning. | FRIEDA: Benchmarking Multi-Step Cartographic Reasoning in VLMs | 2,026 | ICLR 2026 main conference | University of Minnesota Knowledge Computing Lab / Yao-Yi Chiang group, with UC Davis and USC collaborators | https://openreview.net/forum?id=QQCadccQqU | https://github.com/knowledge-computing/FRIEDA | https://huggingface.co/datasets/knowledge-computing/FRIEDA | verified_public_urls | maps; cartography; geospatial reasoning | map understanding; cartographic reasoning; topological reasoning; metric reasoning; directional reasoning; open-ended QA | MLLM | Open-ended cartographic QA benchmark; exact released count should be verified from dataset repository | accuracy; LLM judge; human ratings; task-specific reasoning accuracy | High-priority GSTA benchmark for map/cartographic reasoning; official GitHub README links the Hugging Face dataset release. |
GTR-Bench | Evaluates vision-language models on geo-temporal reasoning over maps, multi-camera video, and moving objects in real-world scenes. | GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models | 2,026 | ICLR 2026 main conference | Tsinghua University geo-temporal vision group (Long Zeng), with SenseTime Research, Peking University, and TUM collaborators | https://openreview.net/forum?id=ecJDyT5qIQ | https://github.com/X-Luffy/GTR-Bench | https://github.com/X-Luffy/GTR-Bench/tree/main/data | verified_repo_data_source_datasets_external | geospatial reasoning; temporal reasoning; maps; video; mobility | geo-temporal reasoning; cross-view reasoning; video understanding; map understanding; object motion reasoning | MLLM | 420 questions and 364 videos | accuracy; multiple-choice accuracy; human comparison | Core benchmark for dynamic geo-temporal VLM reasoning; official repository contains benchmark data structure and requires external source datasets for full use. |
GeoMMBench | Evaluates expert-level multimodal intelligence in geoscience and remote sensing across sensor modalities and domain knowledge tasks. | GeoMMBench and GeoMMAgent: Toward Expert-Level Multimodal Intelligence in Geoscience and Remote Sensing | 2,026 | CVPR 2026 | RIKEN AIP geospatial multimodal group, with Wuhan University, Linkoping University, University of Tokyo, and NUIST collaborators | https://arxiv.org/abs/2604.08896 | https://github.com/NathanZhang/geommagent | https://huggingface.co/datasets/AR-X/GeoMMBench | verified_public_urls | remote sensing; geoscience; Earth observation; GIS | remote sensing VQA; geoscience QA; expert knowledge reasoning; multimodal reasoning | MLLM | 1,053 expert-level image multiple-choice questions | accuracy; multiple-choice accuracy; task/category accuracy | High-priority expert-level remote-sensing/geoscience MLLM benchmark; official repository links the Hugging Face dataset release. |
VIR-Bench | Evaluates MLLMs' geospatial and temporal understanding through travel-video itinerary reconstruction over places and temporal/spatial relations. | VIR-Bench: Evaluating Geospatial and Temporal Understanding of MLLMs via Travel Video Itinerary Reconstruction | 2,026 | AAAI 2026 | Waseda University NLP group, with CyberAgent, AI Shift, and NAIST collaborators | https://ojs.aaai.org/index.php/AAAI/article/view/37938 | https://github.com/nlp-waseda/VIR-Bench | https://soya.infini-cloud.net/share/1302266998c5d047 | verified_public_research_only_data_url | travel; mobility; geospatial reasoning; video | video understanding; itinerary reconstruction; POI reasoning; temporal ordering; spatial relation reasoning | MLLM | 200 Japanese travel videos | node accuracy; edge accuracy; graph similarity; task-specific itinerary metrics | Included as a focused travel-video geo-temporal benchmark; official repository links a research-use dataset download with redistribution restrictions. |
UrbanFeel | Evaluates multimodal models on urban street-view temporal change understanding and subjective urban perception questions. | UrbanFeel: Benchmarking Urban Visual Temporal Reasoning and Perception in Multimodal Large Language Models | 2,026 | ICLR 2026 main conference | Sun Yat-sen University urban visual intelligence group (Weijia Li and Xiang Zhang) | https://mlanthology.org/iclr/2026/he2026iclr-urbanfeel/ | https://github.com/Hejun0915/UrbanFeel | https://huggingface.co/datasets/JunHe0915/UrbanFeel | verified_public_urls | urban science; street-view; temporal change; urban perception | street-view QA; temporal change understanding; visual perception; subjective urban assessment | MLLM | 14.3K visual questions across 11 cities | accuracy; LLM judge; human ratings; task-specific metrics | Useful complement for urban temporal perception; geographic reasoning component is weaker than FRIEDA, GTR-Bench, MapEval, or USTBench. |
TimeSpot | Benchmarks geo-temporal understanding in vision-language models in real-world settings involving place and time cues. | TimeSpot: Benchmarking Geo-Temporal Understanding in Vision-Language Models in Real-World Settings | 2,026 | ICML 2026 | Computational Intelligence and Operations Laboratory (CIOL), SUST/NSU Bangladesh, with QCRI collaborators | https://arxiv.org/abs/2603.06687 | null | https://huggingface.co/datasets/kagnlp/TimeSpot | code_not_linked | geo-temporal reasoning; geolocation; visual scenes | geo-temporal understanding; visual QA; geolocation; temporal reasoning | MLLM | Real-world geo-temporal VLM benchmark; exact public count should be verified from release | accuracy; task-specific geo-temporal metrics | Included as a candidate/near-core geo-temporal VLM benchmark; Hugging Face dataset is public, but the public project/card pages do not link a canonical code repository. Project page: https://timespot-gt.github.io/. |
ERGeoBench | Evaluates multimodal large language models on embodied reasoning and geo-localization in geospatial environments. | ERGeoBench: A Comprehensive Benchmark for Embodied Reasoning and Geo-localization in Multimodal Large Language Models | 2,026 | arXiv 2026 | Beijing University of Posts and Telecommunications networking/AI group, with SJTU, China Mobile, and NTU collaborators | https://arxiv.org/abs/2605.31251 | https://github.com/kaiXuewen/ERGeoBench | https://github.com/kaiXuewen/ERGeoBench/tree/main/data | metadata_release_on_github_full_release_pending | geolocation; embodied reasoning; navigation; visual scenes | embodied reasoning; geo-localization; visual QA; navigation reasoning | MLLM | Benchmark metadata/release pending; exact count should be verified from repository release | accuracy; geolocation error; task-specific reasoning metrics | Included as a candidate benchmark with public arXiv and GitHub evidence; official repository currently exposes metadata/panorama IDs while full release is marked coming soon. |
EarthSpatialBench | Benchmarks spatial reasoning capabilities of multimodal LLMs on Earth imagery. | EarthSpatialBench: Benchmarking Spatial Reasoning Capabilities of Multimodal LLMs on Earth Imagery | 2,026 | arXiv 2026 / OpenReview | University of Florida GeoAI group (Zhe Jiang), with Indiana University collaborators | https://arxiv.org/abs/2602.15918 | null | https://huggingface.co/datasets/anonymous-submission-111/EarthSpatialBench | code_ref_not_linked | Earth observation; remote sensing; spatial reasoning | Earth imagery reasoning; spatial reasoning; visual QA | MLLM | Earth-imagery spatial reasoning benchmark; exact public count not verified | accuracy; task-specific spatial reasoning metrics | Included as a borderline but relevant Earth-imagery MLLM benchmark; Hugging Face dataset is public and references accompanying code, but no canonical code repository URL is linked. |
CityLens | Evaluates multimodal models on city-scale visual understanding and urban reasoning tasks using street-level and/or urban visual data. | CityLens: Benchmarking Multimodal Large Language Models for City-Scale Urban Understanding | 2,026 | ICLR 2026 main conference | Tsinghua University FIB Lab / BNRist urban computing group (Yong Li), with HKUST(GZ) Information Hub and BJTU collaborators | https://mlanthology.org/iclr/2026/liu2026iclr-citylens/ | https://github.com/tsinghua-fib-lab/CityLens | https://huggingface.co/datasets/Tianhui-Liu/CityLens-Data | verified_public_urls | urban science; street-view; city-scale understanding | urban visual QA; city-scale reasoning; scene understanding; spatial reasoning | MLLM | City-scale urban MLLM benchmark; exact public count should be verified from repository/data release | accuracy; task-specific visual QA metrics; LLM judge | Included as an urban-visual complement to CityBench and UrbanFeel; official GitHub README links the public Hugging Face dataset release. |
GeoAgentBench | Evaluates tool-augmented LLM agents on dynamic GIS spatial-analysis workflows with sandboxed tool execution, runtime feedback, parameter accuracy, and map-output verification. | GeoAgentBench: A Dynamic Execution Benchmark for Tool-Augmented Agents in Spatial Analysis | 2,026 | arXiv 2026 | Central South University GeoX Lab / School of Geosciences and Info-Physics, with HNUST collaborator | https://arxiv.org/abs/2604.13888 | https://github.com/GeoX-Lab/GABench | https://github.com/GeoX-Lab/GABench/tree/main/benchmark; https://github.com/GeoX-Lab/GABench/tree/main/dataset | verified_repo_data | GIS; spatial analysis; geospatial analytics; GeoAI | tool-augmented GIS workflow execution; spatial analysis; parameter inference; runtime error recovery; map generation | LLM | 53 spatial analysis tasks; 117 atomic GIS tools; 6 core GIS domains | TAO; TIO; TEM; PEA; VLM Average Score; Eff_macro; Eff_micro | Included as a dynamic execution complement to GeoBenchX and GeoAnalystBench; task construction starts from 50 GeoAnalystBench cases, excludes 10, reconstructs the remaining 40 for atomic tool orchestration, and adds 13 hydrological tasks, with repository data managed through Git LFS and no explicit GitHub license obser... |
GPSBench | A pure-text benchmark for evaluating whether LLMs understand GPS coordinates and real-world geography, spanning geometric coordinate operations such as distance and bearing computation plus coordinate-grounded world-knowledge reasoning. | GPSBench: Do Large Language Models Understand GPS Coordinates? | 2,026 | arXiv 2026 | University of Melbourne NLP and data management group (Jey Han Lau and Jianzhong Qi) | https://arxiv.org/abs/2602.16105 | https://github.com/joey234/gpsbench | https://github.com/joey234/gpsbench/tree/main/data | verified_repo_data | GPS; geospatial reasoning; coordinates; location-based services | GPS coordinate reasoning; distance and bearing computation; coordinate conversion; geolocation knowledge; route and spatial relation reasoning | LLM | 57,800 samples across 17 tasks; Pure GPS and Applied tracks | accuracy; 1-MAPE for numerical computation; task-level aggregate scores | Included as a core pure-text benchmark for intrinsic GPS/geospatial reasoning rather than tool use. Official GitHub repository exposes reproducible code and a data directory; evaluation uses task-specific ground-truth matching and numeric tolerances rather than LLM-as-a-Judge. No explicit GitHub license observed at ver... |
UrbanGeoEval | A city-scale pure-text benchmark designed to disentangle LLM urban factual memory from geospatial reasoning by combining map-derived knowledge questions with in-context reasoning scenarios. | UrbanGeoEval: A City-Scale Benchmark for Evaluating Large Language Models in Geospatial Reasoning | 2,026 | ACL ARR 2026 January submission | Anonymous ACL ARR submission; public OpenReview PDF does not disclose affiliation yet | https://openreview.net/forum?id=WmbvH3wuBT | null | null | paper_verified_code_data_pending | urban science; city-scale geospatial reasoning; maps; POIs; road networks | urban memory; map-based QA; distance and direction reasoning; topology; prediction; planning; navigation; recommendation | LLM | Knowledge Module: 27,000 multiple-choice questions across 9 cities; Reasoning Module: 3,148 human-annotated open-ended tasks | accuracy; formula-based penalties; exact match; Hamming distance; LLM-as-a-Judge rubric scores | Included as a candidate pure-text urban GST benchmark. The paper states dataset and codebase will be released upon acceptance; current public evidence is the OpenReview paper rather than a released repository. |
Compositional GeoQA | A pure-text compositional geospatial QA benchmark generated from OpenStreetMap and Wikidata to evaluate LLM reasoning over spatial constraints and entity constraints. | Automatic Generation of a Compositional QA Benchmark for Geospatial Reasoning | 2,026 | EACL 2026 Student Research Workshop | NAIST geospatial NLP group, with NICT and MatBrain collaborators | https://aclanthology.org/2026.eacl-srw.61/ | https://github.com/NAIST-geo-and-lang/Compositional_GeoQA_Benchmark | https://github.com/NAIST-geo-and-lang/Compositional_GeoQA_Benchmark/tree/main/datasets | verified_repo_data | geospatial QA; OSM; Wikidata; entity grounding; spatial constraints | compositional geographic QA; spatial-entity constraint reasoning; distance thresholds; directional and containment constraints; place-name answer generation | LLM | 5,309 questions generated from 14 question templates | exact string match against gold answer set; accuracy | Included as a compact pure-text diagnostic benchmark for spatial constraints plus entity grounding. Questions are Japanese-only, with English examples/translations mainly for reference. It is EACL SRW rather than a main-conference benchmark, so influence/venue strength should be interpreted conservatively. |
TP-RAG | A pure-text retrieval-augmented travel planning benchmark for evaluating spatiotemporal-aware LLM agents that use POI and trajectory references to generate travel plans. | TP-RAG: Benchmarking Retrieval-Augmented Large Language Model Agents for Spatiotemporal-Aware Travel Planning | 2,025 | EMNLP 2025 main conference | HKUST(GZ) AI Thrust / USAIL urban AI group (Hao Liu and Hui Xiong), with Baidu collaborators | https://aclanthology.org/2025.emnlp-main.626/ | https://github.com/usail-hkust/TP-RAG | https://github.com/usail-hkust/TP-RAG/tree/main/query_benchmark; https://github.com/usail-hkust/TP-RAG/tree/main/agentpedia | verified_repo_data | travel planning; mobility; POIs; trajectory retrieval; spatiotemporal planning | retrieval-augmented itinerary generation; spatiotemporal travel planning; trajectory-level knowledge utilization; query-specific contextualization | LLM | 2,348 real-world travel queries; 85,575 annotated POIs; 18,784 high-quality travel trajectory references | failure rate; repetition rate; distance margin ratio; start time rationality; duration underflow ratio; time buffer ratio; POI popularity; POI relevance; time schedule relevance; LLM-as-a-Judge for selected metrics | Included as an application-level pure-text GST planning benchmark. It complements TripCraft by emphasizing retrieval-augmented, trajectory-aware spatiotemporal rationality rather than only basic plan validity. Full evaluation includes selected LLM-as-a-Judge metrics and requires an evaluator LLM/API setup plus the rele... |
π GSTA: Geo-SpatioTemporal Atlas
A curated benchmark atlas for geospatial-temporal LLM & MLLM research
GSTA is a lightweight, metadata-only registry of benchmarks for geospatial, spatial, temporal, urban, mobility, map, navigation, remote-sensing, and Earth-observation research with language models.
Quick Start Β· At a Glance Β· Schema Β· Contribute
GSTA does not redistribute benchmark samples. Each row is a curated metadata entry pointing to the original paper, code, and data sources when stable public URLs were verified. Please always check the original benchmark license, access policy, and citation requirements before downloading or redistributing linked resources.
β¨ Why GSTA?
Geospatial-temporal evaluation is scattered across urban computing, remote sensing, map reasoning, navigation, GIS, mobility, Earth observation, and multimodal scene understanding. GSTA provides a compact atlas for researchers who need to answer questions like:
- Which benchmarks evaluate geospatial reasoning in LLMs or MLLMs?
- Which datasets are text-only, multimodal, map-based, remote-sensing, or tool-oriented?
- What papers, code repositories, public data URLs, metrics, and caveats should I check before choosing an evaluation suite?
- How does benchmark coverage differ across domains, modalities, and years?
π¦ Dataset at a Glance
benchmark entries |
16 metadata columns |
Paper-linked publication metadata |
LLM + MLLM model interfaces |
| Maps navigation & routing |
Urban mobility & city tasks |
Remote Sensing Earth observation |
Spatiotemporal reasoning & planning |
Current interface coverage
| Model interface | Typical inputs |
|---|---|
MLLM |
images, map snapshots, remote-sensing imagery, video, visual scenes |
LLM |
text, tables, coordinates, trajectories, structured spatiotemporal records |
LLM; MLLM |
mixed text/tool/visual settings |
Included benchmark inventory
GeoBench Β· CityBench Β· UrBench Β· MapEval Β· GEOBench-VLM Β· XLRS-Bench Β· STARK Β· TurnBack Β· TripCraft Β· GeoBenchX Β· GeoAnalystBench Β· USTBench Β· FRIEDA Β· GTR-Bench Β· GeoMMBench Β· VIR-Bench Β· UrbanFeel Β· TimeSpot Β· ERGeoBench Β· EarthSpatialBench Β· CityLens Β· GeoAgentBench Β· GPSBench Β· UrbanGeoEval Β· Compositional GeoQA Β· TP-RAG
π Quick Start
from datasets import load_dataset
# The dataset is a metadata table; each row describes one benchmark.
ds = load_dataset("zhangdw/GSTA", split="train")
print(ds.column_names)
print(ds[0])
Prefer pandas for fast inspection?
import pandas as pd
url = "https://huggingface.co/datasets/zhangdw/GSTA/resolve/main/gsta_benchmarks.csv"
df = pd.read_csv(url)
# Example: find multimodal map / remote-sensing benchmarks
view = df[df["model_type"].str.contains("MLLM", na=False)]
print(view[["benchmark", "domain", "task_type", "metrics"]].head())
ποΈ Dataset File
| File | Description |
|---|---|
gsta_benchmarks.csv |
One benchmark per row. This is the canonical metadata table for the current public snapshot. |
π§ Example Discovery Workflows
Find benchmarks by model interface
llm_only = df[df["model_type"].eq("LLM")]
multimodal = df[df["model_type"].str.contains("MLLM", na=False)]
Find navigation, map, or mobility benchmarks
mask = df["domain"].str.contains("map|navigation|mobility|urban", case=False, na=False)
print(df.loc[mask, ["benchmark", "domain", "task_type", "paper_url"]])
Audit rows that need manual URL follow-up
needs_code = df["code_url"].isna() | df["code_url"].eq("")
needs_data = df["data_url"].isna() | df["data_url"].eq("")
print(df.loc[needs_code | needs_data, ["benchmark", "code_url", "data_url", "url_status", "notes"]])
π§± Schema
| Column | Description |
|---|---|
benchmark |
Benchmark name or acronym; should be unique within GSTA. |
abstract |
One-sentence summary of what the benchmark evaluates. |
paper_title |
Title of the paper that introduced or released the benchmark. |
publication_year |
Publication year of the benchmark paper. |
venue |
Venue and track if needed. |
research_group |
Main organization, lab, or research group behind the benchmark paper, based on public author affiliations and official project/repository ownership. |
paper_url |
Canonical paper URL. |
code_url |
Official code repository URL, if verified. |
data_url |
Official public data or access URL(s), if verified; semicolon-separated for multiple URLs. |
url_status |
URL verification, access, and release-status note, especially for rows with blank code_url cells or non-direct data access. |
domain |
Domain or scenario covered by the benchmark. |
task_type |
Semicolon-separated short task labels. |
model_type |
LLM, MLLM, or LLM; MLLM. |
sample_count |
Core benchmark scale. |
metrics |
Main evaluation metrics. |
notes |
Caveats, release status, count notes, or registry-specific comments. |
π§© Model Type Convention
| Label | Meaning |
|---|---|
LLM |
Text-based language models, including standard LLMs and reasoning-oriented LLMs/LRMs. Coordinates, trajectories, tables, sensor values, and time series still count as LLM if serialized as text or structured text. |
MLLM |
Multimodal large language models with non-text inputs such as images, remote-sensing imagery, map images, videos, visual scenes, or raster data. |
LLM; MLLM |
Benchmarks that evaluate both text-based and multimodal model interfaces. |
β Intended Use
GSTA is designed for:
- literature review and benchmark discovery;
- selecting evaluation suites for geospatial-temporal LLM/MLLM research;
- tracking benchmark coverage across domains, modalities, metrics, and venues;
- maintaining lab-internal or community-facing benchmark inventories;
- identifying gaps in current geospatial-temporal model evaluation.
π€ Contributing a Benchmark
Community contributions are welcome. A high-quality new entry should include:
- a stable benchmark name;
- paper title, year, venue, and canonical paper URL;
- verified official code and data links when available;
- concise domain and task labels;
- model interface label:
LLM,MLLM, orLLM; MLLM; - sample count and primary metrics;
- notes for caveats, missing public releases, version differences, or count discrepancies.
If a benchmark has multiple released subsets or versions, prefer a conservative row-level summary and document the version-specific details in
notes.
π οΈ Maintenance Notes
This public snapshot is maintained as a live CSV registry. Some rows intentionally leave code_url empty when a stable public code repository was not verified, and some data_url entries point to official access instructions rather than direct downloads. url_status records the current follow-up state for those cases. GSTA favors traceability over completeness: uncertain URLs should remain blank until verified against the original release channel.
π Citation
If you use GSTA, please cite this dataset and the original benchmark papers you rely on.
@misc{gsta2026,
title = {GSTA: Geo-SpatioTemporal Atlas},
author = {Dawei Zhang},
year = {2026},
howpublished = {Hugging Face Dataset},
url = {https://huggingface.co/datasets/zhangdw/GSTA}
}
GSTA aims to make geospatial-temporal evaluation easier to discover, compare, and maintain.
- Downloads last month
- 85