Climate Vulnerability Scores for Three Chinese Provinces with LLM Framework

A 2026 study by Aihua Zhang presents a framework for climate vulnerability assessment using Large Language Models and conformal prediction. The 18.6 KB document contains empirical validation results from Guangdong, Sichuan, and Yunnan provinces. It reports performance metrics like a calibration correlation of 0.816 and a root mean square error of 7.8.

Use Cases

• Validate the proposed LLM-conformal prediction framework using reported metrics like calibration correlation (0.816) and actual coverage rate (94.7%).
• Compare Climate Vulnerability Scores (e.g., Guangdong: 47.8, Sichuan: 63.2, Yunnan: 72.6) against external indices like ND-GAIN (r = -0.70) and INFORM (r = 0.77).
• Analyze the performance contribution of LLM integration (11.0% improvement) and conformal prediction (4.7% improvement) from the ablation experiments.
• Assess cross-model consistency between GPT-4 and Claude using the reported validation correlation (r = 0.883, p < 0.001).

Strengths

• Framework validated across three distinct Chinese provinces (Guangdong, Sichuan, Yunnan).
• Reports specific performance metrics: root mean square error of 7.8 and a 16.2% improvement over traditional methods.
• External validation performed against established indices (ND-GAIN, INFORM).

Limitations

• Dataset is a 18.6 KB document describing a framework, not a primary data table with rows and columns.
• Empirical validation is limited to three provinces, offering a narrow geographic scope.
• Underlying raw data used to generate the scores and metrics is not included.

Provenance

Source

Aihua Zhang, published on figshare.

Collection Method

Proposed framework combining LLMs (GPT-4, Claude) with split conformal prediction for uncertainty quantification.

Time Range

Study published in 2026; temporal coverage of underlying assessment data is not specified.

Freshness

Last updated in April 2026.

Geography

Three Chinese provinces: Guangdong, Sichuan, and Yunnan.