GCA Framework: A GCC Countries-Grounded Dataset and Agentic Pipeline for Climate Decision Support

The GCA framework introduces a GCC Countries–Grounded Dataset and Agentic Pipeline for Climate Decision Support, which unifies a curated multimodal dataset with an agentic tool-augmented system to translate heterogeneous scientific and policy evidence into actionable guidance for climate decision-making in the Gulf states. The gist: GCA addresses the gap by aligning Gulf-specific supervision (GCA-DS) with tool-augmented inference for decision-oriented climate queries.

Dataset Curation and Grounding

The framework builds upon GCA-DS, a curated multimodal dataset grounded in the GCC states. This dataset comprises ∼200k question–answer pairs spanning governmental policies and adaptation plans, NGO and international frameworks, academic literature, and event-driven reporting on heatwaves, dust storms, and floods. The textual data (110k) aggregates Gulf-relevant sources including government climate policies and adaptation strategies, NGO and international organization reports (by Gulf state), academic literature on regional climate and resilience, and news/emergency documentation covering recent extremes. The visual-temporal data (90k) is constructed by ingesting multi-source climate and atmospheric products such as CMIP6 HighResMIP models, CAMS forecasts, air-quality reanalyses, and hydrological signals. The visual component involves constructing time-indexed visual artifacts like trends and anomaly plots from standardized meteorological and land-surface variables.

Agentic Pipeline Architecture

The Gulf Climate Agent (GCA) is a tool-augmented language agent designed to answer Gulf-focuse climate queries by combining domain grounding from GCA-DS with structured access to specialized climate analytics tools. The architecture follows an act–observe–reason paradigm where the model selects tools, consumes their outputs, and iteratively refines intermediate hypotheses until it can produce an evidence-grounded response. The GCA framework orchestrates a modular tool pipeline grounded in real-time and historical signals and geospatial processing that produces derived indices and interpretable visualizations.

Tool Suite Integration

The agent utilizes a modular tool suite grouped into six categories to cover Gulf climate objectives. These categories include Remote Sensing and Land Surface (for satellite retrieval and desertification analysis), Biodiversity and Species detection, Web Retrieval and Summarization, Carbon and Sustainability estimation, Air Quality and Atmospheric Composition inquiries, Weather/Rainfall/Hydrology analysis, and Geospatial utility like geocoding. Each tool is exposed to the LLM through a structured function signature requiring typed inputs and outputs, with outputs normalized (units, timestamps) to support multi-step composition.

Benchmarking and Fine-tuning Results

The study evaluates GCA along two axes: domain adaptation via a single parameter-efficient fine-tuning run on the Gulf-focused multimodal dataset and agentic competence via a regression benchmark. Fine-tuning on Qwen2.5-VL 7B yields substantial improvements in tool reliability, showing gains of +32.0 ToolAcc and +31.0 ArgAcc compared to the unadapted baseline. These step-level gains compound into a large end-to-end improvement of +35.9 AnsAcc and +35.1 AnsAcc+I. The error breakdown shows that fine-tuning primarily improves emitting executable tool calls and adhering to schema constraints, which are prerequisites for successful multi-step tool composition. GCA remains competitive with proprietary baselines, exhibiting the lowest error rates among all evaluated models.

Conclusion

The GCA framework contributes a semi-automated curation pipeline producing GCA-DS and a modular tool suite, alongside a benchmarking and finetuning study showing improved tool-use reliability and end-to-end answer accuracy over strong baselines. Future work will expand continual data ingestion, strengthen validation, and add uncertaintyaware and scenario-driven analyses for Gulf climate objectives.

Limitations

GCA has several limitations, including the semi-validated coverage of GCA-DS where automated QA generation can introduce subtle labeling errors. Tool dependence constrains reliability as agent outputs inherit biases and resolution limits from upstream data products. Furthermore, the benchmark scope is necessarily selective as it does not exhaustively cover all climate tasks like full uncertainty quantification. Deployment requires careful monitoring and transparent provenance, especially for high-stakes recommendations.

Appendix A.1 Tool Suite Overview

The tool suite includes functions such as get satellite image to retrieve multispectral satellite imagery and calculate ndvi to quantify vegetation condition. It also includes desertification analysis to compare two images and return land-degradation indicators. Other tools cover biodiversity detection, web retrieval and summarization, carbon footprint calculation, air quality inquiry/prediction/analysis, pollen forecasting, weather inquiry/forecast/analysis, and geospatial utility like geocode mapping.

Table 4: Tool Suite Details

Table 4 details the GCA tool suite, showing each tool is exposed to the LLM with typed inputs/outputs where outputs are normalized to support multi-step composition.

Table 2: Percentages of tool-use error types on GCA

Table 2 summarizes tool-use error rates, showing Format Err., Arg. Err., and N/A errors for various models.

Table 3: Tool-use benchmark results on GCA

Table 3 reports the main benchmark results, showing that GCA is best or competitive across all metrics in both Step-by-step mode and End-to-End mode.

Figure 4: Gulf Climate Agent (GCA) framework

Figure 4 summarizes the GCA framework, showing the flow from Retrieval-Grounded Keyword Expansion through to the final Response.

Figure 3: Distribution of GCA-DS across question types and categories

Figure 3 shows the distribution of GCA-DS across question types and categories, indicating a high volume of OpenEnded and TrueFalse items.

The GCA framework successfully bridges general-purpose LLM reasoning with specialized climate tools and datasets tailored to the Gulf by aligning Gulf-specific supervision with tool-augmented inference.

1. The GCA framework can perform grounded, interpretable climate decision support by unifying a Gulf-focused multimodal dataset (GCA-DS) with an agentic tool pipeline that produces derived indices and visualizations from real-time and historical signals.

2. The system can execute multi-step workflows typical of Gulf climate analysis, such as geocode mapping followed by temporal weather analysis, to produce evidence-grounded responses rather than relying solely on parametric knowledge.

3. Fine-tuning the model on the Gulffocused multimodal dataset improves tool competence, specifically enhancing tool selection accuracy and argument name accuracy, leading to a substantial improvement in end-to-end answer accuracy.

4. The system can perform complex data synthesis by using a Temporal plot builder and then synthesizing QA items in MCQ, open-ended, and true/false formats based on anomaly detection, forecasting, imputation, or reasoning QA across visual-temporal data.

5. By utilizing the modular tool suite—covering remote sensing, air quality/health indices (like AQI), weather forecasting (including precipitation), and hydrology (river discharge)—the agent can provide comprehensive analysis for hazards like heatwaves, dust storms, and floods.

Climate decision-making in the GCC states increasingly demands systems that can translate heterogeneous scientific and policy evidence into actionable guidance, yet general-purpose large language models (LLMs) remain weak both in region-specific climate knowledge and grounded interaction with geospatial and forecasting tools. We present the GCA framework, which unifies (i) GCA-DS, a curated multimodal dataset grounded in the GCC states, and (ii) Gulf Climate Agent (GCA), a tool-augmented agent for climate analysis. GCA-DS comprises 200k question--answer pairs spanning governmental policies and adaptation plans, NGO and international frameworks, academic literature, and event-driven reporting on heatwaves, dust storms, and floods, complemented with remote-sensing inputs that couple imagery with textual evidence. Building on this foundation, the GCA agent orchestrates a modular tool pipeline grounded in real-time and historical signals and geospatial processing that produces derived indices and interpretable visualizations. Finally, we benchmark open and proprietary LLMs on climate tasks in the GCC states and show that domain fine-tuning and tool integration substantially improve reliability over general-purpose baselines.