Climate answers, grounded in the Gulf.
وكيل مناخ الخليج
Gulf Climate Agent (GCA) is a tool-augmented language model built for the six GCC states. It reads policy documents, satellite imagery and reanalysis data, calls specialised climate tools, and returns answers you can trace back to evidence.
General models know climate. They don't know the Gulf.
General-purpose LLMs are weak on region-specific climate knowledge, numerical reasoning over climate plots, and jurisdiction-specific policy. GCA closes that gap with Gulf-grounded supervision and direct access to the data products decision-makers rely on.
Extreme heat
Heat forecasting, temperature anomalies and UV exposure at city level.
weather_analysis · uv_index_forecastDust storms
Particulate trends, exceedances and air-quality forecasts during dust events.
aqi_analysis · aqi_predictionFlash floods
Rainfall extremes and river discharge from GloFAS hydrological signals.
rain_analysis · river_discharge_checkLand & coast
Desertification and moisture change from two-date satellite comparisons.
calculate_ndvi · desertification_analysisAct, observe, reason. Then answer.
GCA keeps a trajectory of tool calls and observations. At each step it infers what kind of question it faces, calls the smallest set of tools that can answer it, and checks the result before going further.
Infer intent
Classify the dominant need behind the question before touching any tool.
Route to tools
Narrow to a minimal tool family and emit a typed call with arguments.
Execute & validate
Run the call, then check schema, units and timestamps on the returned observation.
Chain & synthesise
Compose more tools if needed, then write a response grounded in their outputs.
A multimodal climate dataset built from Gulf sources.
About 110k text items from government policy, NGO reports, academic papers and event news, plus 90k visual-temporal items rendered from reanalysis, atmospheric composition and hydrology products for cities across all six states.
Two curation pipelines
Automated extraction and synthesis, with human-in-the-loop verification on a subset.
- Keyword expansioncos-sim dedup < τ
- Web retrieval agentReAct loop
- Document parsingHTML · PDF
- ChunkingL=512, S=384
- Fact inductionatomic claims
- QA synthesisMCQ · T/F · open
- Multi-source ingestERA5 · CAMS · GloFAS
- Geocodingcity → (φ, λ)
- Spatial index0.05°–0.1° grids
- Unit normalisationcanonical CSV
- Plot builderΔ = 90 days, 10 yrs
- Visual QAanomaly · forecast
Composition
201,410 items across three question formats and twelve topics.
| Dataset | Size | Automated | Validated | Multimodal | Region | Remote sensing | Topics |
|---|---|---|---|---|---|---|---|
| Climate Crisis QA (2024) | 19,241 | ✓ | ✗ | ✗ | General | ✗ | 5 |
| Pirá 2.0 (2024) | 2,250 | ✗ | ✓ | ✗ | General | ✗ | 4 |
| Climate-FEVER (2020) | 1,535 | ✗ | ✓ | ✗ | General | ✗ | 3 |
| CPIQA (2025) | 54,612 | ✓ | Partial | ✓ | General | ✗ | 5 |
| ELLE (2025) | 1,130 | ✗ | ✓ | ✗ | General | ✗ | 6 |
| ClimaQA-Gold (2025) | 566 | ✓ | ✓ | ✗ | General | ✗ | 5 |
| ClimaQA-Silver (2025) | 3,000 | ✓ | ✗ | ✗ | General | ✗ | 4 |
| MMClima (2025) | 104,902 | ✓ | ✓ | ✓ | General | ✗ | 5 |
| GCA-DS (ours) | 201,410 | ✓ | Partial | ✓ | Gulf-specific | ✓ | 12 |
Comparison of climate QA datasets (paper, Table 1). "Partial" means limited human validation such as expert-guided prompting or subset review.
Every tool has a typed signature.
Outputs are normalised for units, timestamps and uncertainty, so the agent can chain calls such as geocode → weather_analysis → summarize without ad hoc parsing.
More reliable tool use than frontier models.
One LoRA fine-tuning run on GCA-DS turns Qwen2.5-VL 7B from the weakest model in the tool-use benchmark into the strongest on five of six metrics, ahead of GPT-5, Claude 4.5 Sonnet and Gemini 2.5 Pro.
Tool-use benchmark
Gain from fine-tuning
GCA vs. its unadapted Qwen2.5-VL 7B base.
Tool-call errors
Share of calls, lower is better.
- Backbone
- Qwen2.5-VL 7B
- Method
- LoRA r=16, α=16
- Optimiser
- AdamW 8-bit, cosine
- Learning rate
- 5×10⁻⁵ (emb. 1×10⁻⁵)
- Benchmark
- 91 annotated questions
Known limitations
Partial validation
Only a subset of GCA-DS is human-verified. Automated QA generation can carry label or grounding errors.
Tool dependence
Outputs inherit biases, resolution limits and data gaps from upstream products, and tool failures can propagate.
Selective benchmark
The suite targets Gulf tool workflows, not full uncertainty quantification, attribution or long-horizon scenarios.
Expert oversight
High-stakes use needs monitoring, transparent provenance and review by domain experts.
GCA Framework: A GCC Countries–Grounded Dataset and Agentic Pipeline for Climate Decision Support
1 Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), UAE · 2 University of Missouri, USA · 3 Australian National University · 4 Linköping University, Sweden
Climate decision-making in the GCC states increasingly demands systems that can translate heterogeneous scientific and policy evidence into actionable guidance. We present the GCA framework, which unifies GCA-DS, a curated multimodal dataset of ~200k question–answer pairs grounded in the GCC states, and Gulf Climate Agent, a tool-augmented agent for climate analysis. We benchmark open and proprietary LLMs on Gulf climate tasks and show that domain fine-tuning and tool integration substantially improve reliability over general-purpose baselines.
@article{sheikh2026gca,
title = {GCA Framework: A GCC Countries--Grounded Dataset
and Agentic Pipeline for Climate Decision Support},
author = {Sheikh, Muhammad Umer and Shehzad, Khawar and
Khan, Salman and Khan, Fahad Shahbaz and
Khan, Muhammad Haris},
journal = {arXiv preprint arXiv:2604.12306},
year = {2026}
}