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

Muhammad Umer Sheikh, Khawar Shehzad, Salman Khan, Fahad Shahbaz Khan, Muhammad Haris Khan · 2026

Read GCA Framework: A GCC Countries–Grounded Dataset and Agentic Pipeline for Climate Decision Support by Muhammad Umer Sheikh, Khawar Shehzad, Salman Khan, Fahad Shahbaz Khan, Muhammad Haris Khan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-12306.txt.

Introduction

Sheikh and four colleagues present the GCA framework for climate decision support in the Gulf states, pairing a curated dataset with a tool-using agent. GCA-DS holds 200,000 question-answer pairs across policies, adaptation plans, frameworks, literature, and event reporting on heatwaves, dust storms, and floods, plus remote-sensing inputs. The Gulf Climate Agent orchestrates real-time and historical signals with geospatial processing into indices and visualizations. Domain fine-tuning and tool use beat general-purpose baselines on reliability.

Catalog record

CREATORMuhammad Umer Sheikh, Khawar Shehzad, Salman Khan, Fahad Shahbaz Khan, Muhammad Haris Khan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35638
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.12306

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GCA Framework: A GCC Countries–Grounded Dataset and Agentic Pipeline for Climate Decision Support Muhammad Umer Sheikh, Khawar Shehzad, Salman Khan, Fahad Shahbaz Khan, Muhammad Haris Khan Abstract 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 ∼ \sim 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. Figure 1: Overview of Gulf Climate Agent (GCA) Framework. We curate a Gulf-focused multimodal QA dataset, GCA-DS and fine-tune a tool-augmented LLM that routes user queries to specialized climate tools to produce grounded, interpretable outputs 1 Introduction Climate change is among the most consequential challenges for societies worldwide, yet its impacts and policy responses are profoundly shaped by regional context ( 18 ) . The Gulf region, encompassing the United Arab Emirates, Saudi Arabia, Qatar, Kuwait, Oman, and Bahrain , faces a unique constellation of climate hazards: extreme heat, dust storms, flash floods and rapid coastal erosion. These challenges demand decision support systems that can distill complex geophysical data and policy documents into actionable insights. While recent advances in large language models (LLMs) and vision–language models (VLMs) have made it possible to access and summarize climate information at scale, general-purpose models often fall short on fine-grained tasks such as numerical reasoning over climate visualizations (trend/anomaly plots), tool-mediated retrieval of spatiotemporal variables (AQI, rainfall, discharge), and jurisdiction-specific policy interpretation. High-stakes decisions in infrastructure and policy require numerically precise, source-grounded answers that adhere to domain conventions ( 3 ; 28 ) , yet many models struggle with figure understanding and visual reasoning ( 27 ; 29 ) . Existing climate QA benchmarks either rely on small expert-curated corpora or automatically generated datasets that suffer from noise and weak validation, and they seldom incorporate multimodal evidence or region-specific context ( 26 ) . Within climate NLP, recent benchmarks and datasets have begun to systematize evaluation and introduce multimodal supervision. For example, 3 studies the adequacy of LLM responses to climate questions under a fine-grained evaluation framework; 26 propose an automated benchmark creation pipeline (ClimaGen) and expert-in-the-loop evaluation; CPIQA introduces a figure-grounded climate QA benchmark from scientific articles, targeting retrieval-augmented setups ( 30 ) . Despite these advances, to our knowledge there is no publicly available resource that (i) is explicitly Gulf-grounded across policy, hazards, and geospatial evidence at large scale, and (ii) is paired with an agentic tool-augmented pipeline designed for Gulf climate objectives. We introduce Gulf Climate Agent (GCA) , a framework that bridges general-purpose LLM reasoning with specialized climate tools and datasets tailored to the Gulf (see Figure 1 ). Semi-automated data-generation framework. We construct a semi-automated Gulf-specific dataset comprising approximately 200k question–answer pairs. The dataset is sourced from government policies and adaptation strategies, NGO reports and international frameworks, academic papers on climate and sustainability, news articles describing recent heatwaves, dust storms and floods, and geospatial/remote-sensing resources (e.g., Sentinel-2 imagery and Google Earth Engine). Our data generation pipeline combines automated extraction and synthesis with human-in-the-loop verification to enable scale while maintaining reliability, grounding responses in authoritative Gulf sources rather than generic internet text. Agentic pipeline with climate-specific tools. We develop an agentic pipeline that orchestrates LLM reasoning with climate-specific tool suites. Our modular architecture links the LLM to specialized modules, including heat forecasting, flood-risk prediction, carbon-footprint estimation, and coastal-erosion analysis, augmented with geospatial processing to produce interpretable maps and trends (heat maps, flood indices, air-quality trajectories, shoreline-change profiles). General-purpose retrieval and retrieval-augmented generation (RAG) components complement these modules, enabling multi-step reasoning that integrates evidence from documents and structured data. Benchmarking and fine-tuning. We fine-tune and benchmark both open and proprietary LLMs (e.g., GPT-family models, Claude, Qwen) on Gulf-centric tasks, including climate question answering, policy summarisation, tool-invocation accuracy, and geospatial reasoning. Our benchmark suite includes manually annotated 91 questions and evaluation metrics for factuality, numerical precision, and tool-use reliability. Fine-tuning on the curated dataset yields domain-specialized models that improve over general-purpose baselines, highlighting the benefits of regional adaptation and tool integration. 2 Related Work Climate QA and Multimodal Climate Benchmarks. Climate QA has evolved from machine reading over curated documents to evaluating foundation models and retrieval-augmented generation (RAG) on technical sources. Early systems such as Climate Bot ( 32 ) demonstrated document-grounded climate QA via CCMRC. More recent benchmarks broaden scope and difficulty: ClimaQA ( 26 ) builds expert-in-the-loop evaluation sets from graduate-level textbooks, CPIQA ( 30 ) introduces figure-grounded climate QA from scientific articles for RAG settings, and MMClima ( 1 ) expands multimodal coverage with expert-validated QA over figures and text. Complementarily, ClimateIQA ( 4 ) targets meteorology-style visual reasoning over heatmaps. Despite these advances, existing resources are largely global and rarely couple regional governance documents with city-level spatiotemporal evidence, which is central to Gulf decision contexts. Dataset Size Automated Validated Multimodal Region Remote Sensing Topic Covered 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 Table 1: Comparison of climate- and environment-focused QA datasets. “Partial” denotes limited human validation (e.g., expert-guided prompting or subset review). Topic Covered reports the number of represented topic types. Evaluating Climate Knowledge and Scientific Reasoning. Several studies assess LLM climate competence and scientific reliability, highlighting gaps between fluent generation and faithful, source-grounded responses ( 3 ; 37 ; 22 ) . In parallel, general scientific QA benchmarks (e.g., ScienceQA, SciQAG, SciQA) provide methodology for multimodal supervision and evaluation design, but remain domain-general and do not test climate-specific tool use or geospatial reasoning ( 25 ; 33 ; 2 ) . Related work on chart and plot QA further shows persistent failure modes in numerical extraction and reasoning over scientific visualizations ( 27 ; 29 ) , motivating tool-mediated computation for climate settings. Agentic Tool Use for Climate Workflows. Agentic systems increasingly operationalize tool-augmented reasoning for complex tasks. ClimateAgent ( 21 ) proposes a climate workflow agent and benchmark, while general tool-learning benchmarks (e.g., AgentBench, ToolLLM/ToolBench, ToolPlanner, StableToolBench) study planning, tool invocation, and evaluation stability across domains ( 24 ; 31 ; 35 ; 16

Read the complete text